Most teams building retrieval augmented generation (RAG) systems for high stakes classification make the same architectural bet: Route every ambiguous case straight to the language model and trust the retrieved context to sort it out. This works fine in a demo. It falls apart the moment the system has to survive an audit, a regulator, or a compliance officer asking why a specific decision was made six months ago.

I have spent the last year building RAG based classification systems in regulated enterprise settings, where the cost of a wrong answer is not a bad chatbot reply. A decision has to hold up to scrutiny long after the model produced it. This environment forces a different design philosophy than most AI engineering content assumes.

Here is what changes when you cannot afford to be probabilistic about everything, and how a cascade architecture solves it.

The invisible cost of an all LLM pipeline

The appeal of routing everything through a large language model (LLM) is obvious: Fewer moving parts, faster iteration, the model handles unanticipated edge cases. The problem shows up later, in three places.

First, auditability. "The model decided based on retrieved context" is not an acceptable answer. You need a decision path a human can reconstruct without rerunning inference and hoping for the same output.

Second, cost at scale. If your system processes tens of thousands of cases a day and every one hits an LLM call with several retrieved documents in context, your inference bill and latency both scale with volume in a way that rule based logic does not.

Third, and least discussed, model drift on the easy cases. LLMs are excellent at nuanced judgment calls. They are inconsistent, in ways that are hard to detect, on cases that should have a deterministic answer. A clear structured match against known criteria should never depend on a language model's mood.

The cascade approach

The fix: Stop treating the LLM as the front line and start treating it as the escalation path. In practice this means a three stage pipeline.

Stage one is deterministic. Exact matches, structured field comparisons, and anything with a clear rule get resolved here with no model call at all. This stage should clear the majority of volume, often more than half depending on your data quality, and every decision is fully explainable because it is a lookup, not an inference.

Stage two is where retrieval earns its keep. For cases that survive stage one — and I mean survive as in they were not clearly resolved — you build a retrieval layer that pulls the specific evidence relevant to the ambiguity: Prior reviewer decisions on similar cases, contextual documents that explain an apparent conflict, or historical precedent that clarifies an edge case. The retrieval step matters more than the generation step here. If you retrieve the wrong context, even the best language model in the world will produce a confident, well reasoned, wrong answer.

Stage three is the LLM call, and it should only see the residue that stages one and two could not resolve. This is the part people skip when they design their first version, and it is the single biggest lever for both cost and quality. In one system I worked on, routing only the genuinely ambiguous 10 to 15% of cases to the LLM cut inference cost by roughly 6X compared to an all LLM baseline, while improving consistency on the deterministic majority to effectively perfect.

Designing the prompt for asymmetric risk

Once a case reaches the LLM stage, most teams default to a neutral prompt: "Assess whether this case should be approved or flagged." That framing is wrong for high stakes classification because the cost of the two error types is not symmetric. Missing something that genuinely needed attention can mean real harm downstream. Incorrectly flagging something that was fine costs a reviewer's time and a delay. Those two outcomes are rarely equally bad, yet a neutral prompt asks the model to treat them as if they were.

An asymmetric risk prompt makes that tradeoff explicit to the model rather than letting it guess at your risk tolerance. Concretely, this means instructing the model to treat uncertainty as a reason to escalate rather than clear, providing calibrated examples of both error types with their consequences spelled out, and asking for a confidence score alongside the classification rather than a binary answer. The confidence score becomes your second cascade point: Anything below a certain threshold goes to a human reviewer instead of being auto resolved, no matter what the model's classification says.

This sounds like a small prompt engineering detail. In practice it is the difference between a system that reduces reviewer workload and one that quietly increases risk while looking like it is working.

Evaluating a system like this properly

Standard RAG evaluation metrics were not built with this use case in mind, and using them without adaptation will give you a false sense of confidence. A few adjustments that matter.

Retrieval quality needs to be measured separately from final classification accuracy. A system can have excellent retrieval ranking scores and still make bad final decisions if the generation step misweights the evidence. Track them independently.

Your evaluation set needs deliberate oversampling of the cases that reach stage three, since that is where your system's judgment actually gets tested. If your eval set mirrors your production distribution, it will be dominated by the deterministic cases your cascade already handles well, and you will be blind to exactly the failures that matter most.

LLM as judge evaluation works for this domain but only if the judge prompt encodes the same asymmetric risk framing as your production prompt. A judge that treats both error types equally will systematically favor the wrong tradeoff when you are tuning your system.

Finally, build a feedback loop from confirmed outcomes back into your retrieval corpus. When a human reviewer overturns a model decision, that case and its correct resolution should become retrievable context for future similar cases. Without this, your system's handling of ambiguous cases never improves, it just keeps making the same category of mistake at the same rate.

The broader lesson

The instinct to reach for the most capable model for every decision is understandable, but in domains where wrong answers have real consequences, the more valuable engineering work is deciding what should never touch the model at all. Cascade architecture is not a workaround for LLM limitations. It is what a mature RAG system looks like once you have actually had to defend its decisions to someone whose job is to find the flaw in your logic.

If you are building AI systems for any regulated or high stakes domain, the question worth asking before you write a single prompt is not "How do I get the model to handle this well." It is "Which parts of this decision should never have been the model's job in the first place."

Vineet Vijay is a Lead AI and machine learning engineer.

DeepSeek's V4 Flash has topped model leaderboards and been hailed by developers as a "total monster" since its rollout. But in real-world testing, it completed just 53.8% of a batch of complex agent tasks.

Composio ran the model through eight different agent harnesses, including Claude Code, Codex, and OpenCode, on 30 deliberately difficult, multi-step tasks spanning live tools like Gmail, GitHub, Slack, and Google Sheets. Of 240 total runs, 129 passed — and only six of the 30 workflows were completed successfully by every harness tested.

The gap illustrates why orchestration, not raw model capability, may decide whether the model succeeds in enterprise settings: the same model produced substantially different results depending on the harness, tool configuration, caching behavior, retries, and provider stack it ran on.

DeepSeek said it will be hiking the prices for V4 Flash and Pro, models that have quickly become favorites among developers building coding assistants and agents.

While it seems the move might undercut its very appeal — strikingly capable models at ultra-low pricing that frontier providers simply can't match — it also moves the story beyond the now-clichéd "cheap Chinese model" narrative, as early use cases emerge and enterprises figure out where different models fit into their tech stacks and what workflows they should be aimed at.

"Insane" adoption numbers as DeepSeek flips the cost structure

DeepSeek rolled out V4 Flash to public beta July 31, and made V4 Pro generally available on August 13. The 284-billion-parameter Flash is built for volume and speed, the 1.6 trillion-parameter Pro for more complex workflows.

Both models have flexible reasoning capabilities (low, high, max) and ‘thinking modes’ applying chain-of-thought (CoT) reasoning to improve answer accuracy.

Users were immediately impressed by Flash’s capabilities. It has dominated OpenRouter's usage leaderboard since its rollout, currently the most-used model on the platform by weekly token volume.

“The adoption numbers of the initial DeepSeek V4 Flash were insane,” ML researcher Nathan Lambert posted to X, adding that the new version "scored the same as GLM 5.2," making it a "total monster" that will be used extensively.

DeepSeek switching the cost model adds an interesting dimension.

V4 API rates are going up by as much as 1,100% depending on the model, token type and time of use. The new pricing structure:

• Flash will be 22 cents per million input tokens and 66 cents per million output tokens off-peak; and 44 cents per million input tokens and $1.32 per million output tokens at peak. This represents a 57% to 371% increase.

• Pro will be 66 cents per million input tokens and $1.98 per million output tokens off-peak; and $1.32 per million input tokens and $3.96 per million output tokens at peak. This shows a 51% to 355% jump.

• Cache hits, meanwhile (when models reuse prompts rather than starting from scratch), are going up between 52% and 1,100%.

DeepSeek says offering 50% lower off-peak usage is intended to encourage “more flexible workload scheduling." Seventeen of every 24 hours stay at half price, and the new structure actually prices the company's home market the highest.

"This is not a simple price rise," said Sanchit vir Gogia of Greyhound Research. "It is a pricing architecture that makes the timing of inference an economic variable."

Work that can wait — such as batch evaluation, synthetic-data generation, and overnight development runs — moves into the cheap hours; interactive agents and live operations cannot. Gogia said irritation among developers and enterprises is genuine and vocal, and that DeepSeek's past low pricing doesn't obligate it to stay cheap forever.

At first glance, it does look like a “suicidal move from a platform still looking for credibility against more established AI model vendors,” said tech analyst Carmi Levy. The increases will certainly eat into DeepSeek's price advantage and force customers to weigh concerns around the company's Chinese origins more heavily.

Still, DeepSeek remains far cheaper by all pricing measures relative to competing models from OpenAI, Anthropic, Google, Cohere, xAI, and others, he said.

So, while the move will force DeepSeek to emphasize performance and security over cost, it hardly wipes out its already-notable price-performance advantage, and still gives customers ample wiggle room to justify its use for specific workloads, Levy said. The math will just have to be more tightly calculated.

“The advantage will likely erode over time as DeepSeek inevitably continues to align pricing with market realities, but for now it's still easy to make the business case,” Levy said.

Where can DeepSeek Flash fit into enterprise environments?

Adoption inside enterprises remains an open question due to cost, capability, reliability, data governance, security, and other factors.

One use case is batch processing, Levy said. This kind of work is typically routine and repetitive rather than demanding top-tier intelligence, so it makes sense to use a cheaper, more efficient model. “It’s a high-performance inference engine that enterprises can consider using for point solution workloads rather than as a wholesale replacement for the incumbent offerings,” Levy said.

Partial adoption will likely involve isolated, non-sensitive workloads with clearly-defined success metrics, strict oversight, permissions controls, and fallback models in case of failures, he noted. Broader deployment will require DeepSeek and its hosting partners to demonstrate strong reliability, security, privacy, auditability, and deployment options. As it adjusts price structures based on demand, DeepSeek also must retain a large enough price-performance advantage to justify any risk, he said.

Expect unsanctioned, smaller-scale use in backroom labs and contained test environments as IT teams get familiar with the new model and figure out when and how to bring it to senior leadership for budget approval.

“DeepSeek has built a well-earned reputation as a global disruptor,” he said, “and it’s clear that its march to broader enterprise adoption will continue to gather momentum.”

Testing DeepSeek in multi-tool workflows

While many use cases are still in the experimental phases, Meta software engineer Naman Ahuja offers one that could translate directly into enterprise environments. In a project unrelated to his employer, he built a home-automation agent with DeepSeek V4 Flash to explore how a lower-cost model performs as the reasoning/orchestration layer for a real multi-tool workflow.

When he leaves home, an agent coordinates several actions across otherwise separate systems: Such as setting a thermostat to “away” to reduce unnecessary energy use, arming a Ring security system, closing and locking doors.

“What interested me was not simply whether the model could understand a command, but whether it could translate intent into a sequence of actions across multiple tools where reliability matters,” Ahuja said.

The biggest lesson was that once a model can take actions, reliability matters as much as intelligence. The system needs structured tool outputs, verification that actions actually succeeded, retry/failure handling, and clear boundaries around what the model is allowed to do.

In the case of enterprise, “the architecture is similar.” Home devices change to ticketing systems, databases, CRM platforms, or infrastructure APIs. The most useful agents will likely orchestrate repetitive workflows across multiple systems, with scoped permissions, auditability, observability, and human approval for higher-risk actions.

“Many valuable AI agents will not be chatbots; they will be background agents coordinating APIs, infrastructure, and business systems in response to events,” he said.

Enterprises need tangible use cases

But Flash's API is still in public beta, Gogia pointed out, and there is not yet an evidence trail of settled enterprise adoption, real-world deployments, and named customers.

“The benchmark story is looser than its retelling, the portfolio story is newer than it looks, and the economics have moved into the system around the model,” he said. Developer mainstreaming is proven; enterprise standardization is not. “The model is mainstream by traffic and still unproven by contract.”

DeepSeek's own integration guidance is an important consideration, he said: Its documentation for at least one popular agent environment states that built-in V4 entries are not sufficient for reliable operation without compatibility overrides.

“Which is a vendor telling the market, accurately, that benchmark performance is not a proxy for production readiness,” Gogia said. “A model can score beautifully and still misbehave once tools, credentials, and state enter the room.”

The serving layer behaves no differently: the same open weights run by different hosts show visible differences in throughput and uptime. "Choosing Flash therefore answers one procurement question and opens three more: Who serves it, where it runs, and which controls surround it," Gogia said.

Prepare for a multi-model future

DeepSeek offers a nuanced case for a multi-model future. V4 Flash is being deployed as the high-volume worker inside diverse estates, Gogia noted: It handles routine generation, retrieval, and background automation, while more difficult or sensitive tasks go elsewhere.

“The question is whether its performance is sufficient for the real-world workflows enterprises actually run, not whether it tops every benchmark,” he said. Enterprises must determine which combination of model, harness, and provider completes the work safely at the lowest cost.

Adam Dalloul, CEO and founder of EmpirioLabs AI, pointed out that bigger isn’t always better; workflows should be task-dependent. For example, his team at EmpirioLabs AI — which hosts 100-plus models on one API, including DeepSeek V4 Flash — were recently working on translating its site into different languages, and there was no need for a large model like GPT 5.6 Sol or Opus 5 to complete the task.

“This is where subagents come in handy,” he said.

His recommended approach: Spawn cheaper subagents and adapt per task. For instance, use Flash variants for day-to-day work, and Pro variants when you need something more powerful. “It depends on the nature of your application.”

Many companies are pivoting towards their own internal benchmarks to route models effectively, Dalloul noted. For example, his team has a workflow that puts a model through various gates and instructions. This helps them identify the model with the speed and accuracy required for the task.

In another example, one of his enterprise clients exclusively wanted access to DeepSeek V4 Flash. They had tested a variety of models and V4 Flash was the only one that met their criteria for speed, cost, and an “appropriate intelligence threshold.”

Meta’s Ahuja agreed that smaller, more efficient models can handle frequent, well-defined agentic tasks, while more expensive frontier models can be reserved for “ambiguous, difficult, or higher-risk decisions.” The relevant metric increasingly becomes cost per successfully completed workflow rather than simply cost per token.

The trade-off, however, is that cheap inference does not automatically mean cheap or safe automation, he said. Once an AI system can take actions, reliability, verification, permissions, failure handling, and security become much more important.

“A failed text response is inconvenient; a failed action in an operational workflow can have real consequences,” Ahuja said.

There is a step in the development process for large language model (LLM)-assisted tooling that most teams skip because it's tedious, time-consuming, and doesn't produce results visible to end users: Verifying that what the model is saying is actually correct. Not fluent, not coherent, not topically relevant — correct in the sense of accurately identifying the right answer to the specific problem the tool was built to solve.

The gap between "this output sounds right to me" and "this output is verifiably correct" is where most LLM-assisted enterprise tools fail quietly. They pass internal review because the output sounds right. They fail in production because those people weren't reviewing against ground truth — they were reviewing against their intuition about what a good answer looks like.

This distinction matters more as LLM-assisted tools move from productivity accessories to components that influence real business decisions. If your AI-assisted tool is shaping how an analyst investigates a data quality issue, how a compliance reviewer decides whether to escalate a flagged record, or how an operations team triages a validation failure — the accuracy of its output has real consequences. "Seems reasonable" is not an adequate evaluation standard for that.

What qualitative evaluation actually catches

The standard evaluation approach for LLM output in enterprise tooling is qualitative: A sample of outputs is reviewed by someone with domain knowledge, judged against a mental model of what a good answer looks like, and the prompt is adjusted if too many outputs seem off.

This catches a specific class of problems: Outputs that are obviously wrong, poorly formatted, or off-topic. These are real issues worth catching. They're also the easy ones.

What qualitative evaluation consistently misses is the class of outputs that are wrong in ways that are difficult to see without checking against something external. An explanation that confidently identifies the wrong root cause, in language that sounds authoritative, based on reasoning that sounds plausible — this passes qualitative review. It fails the moment someone with the right context checks it against what actually happened.

In a system whose value proposition depends on accuracy, "sounds plausible" is not the same as "correct." The two can diverge significantly, and qualitative review won't tell you when they have.

What an actual eval harness looks like

The alternative is building an evaluation harness that scores model output against labeled ground truth — a set of cases where the correct answer is known, against which you can measure accuracy rather than coherence.

I built this while developing a root-cause explainer for data migration drift: A tool that takes a detected drift event and generates a ranked explanation of what most likely caused it. The first prototype produced fluent, specific-sounding explanations that passed qualitative review. When I tested it against cases where I already knew the root cause, the explanation was wrong often enough to matter.

The eval harness I built works in three parts.

First, a synthetic ground truth dataset: Cases where the correct answer is known by construction. This meant introducing specific, controlled causes into a test pipeline — schema changes, transformation logic bugs, source system behavioral shifts — recording exactly what I introduced, and running the model against the resulting drift events. The correct answer for each case was the cause I had deliberately introduced.

Getting the synthetic scenarios realistic enough to be useful required more care than I expected. Early versions were too clean — the drift signal was obvious in ways that real production drift events aren't. Adding realistic noise, overlapping signals, and cases where multiple plausible causes were present simultaneously was what made the synthetic set actually predictive of real-world performance.

Second, a scoring function that evaluates ranked output. Binary correct/incorrect isn't sufficient when the model produces a ranked list of likely causes rather than a single answer. An explanation that correctly identifies the root cause as the third most likely candidate is meaningfully different from one that identifies it as the most likely. The scoring function evaluated two dimensions: Presence — did the correct answer appear in the output at all — and rank — how prominently was it featured relative to incorrect candidates. These were combined into a weighted score that rewarded both finding the right answer and ranking it appropriately.

Third, systematic evaluation across the full synthetic dataset rather than spot-checking. Running the harness across the complete set reveals patterns that spot-checking misses: Which categories of problem the model handles reliably, which it consistently gets wrong, and which combinations of signals produce the highest rate of confident incorrect explanations.

What the evaluation revealed

The results were more informative than any qualitative review could have been.

Schema change scenarios scored well — the model was reliable at identifying upstream schema changes when the evidence was present and distinctive. Transformation logic bugs were harder — the model consistently identified the right general category but misattributed the specific change that caused the problem, particularly when multiple changes had been made close together. Overlapping-signal scenarios were the hardest — cases where two different causes occurred close in time produced the highest rate of confidently wrong explanations.

That last finding is the one that qualitative review would never have surfaced. The model's expressed confidence didn't correlate with its accuracy — it was most confident in the cases where it was most wrong. Without the eval harness measuring against ground truth, that pattern would have been invisible.

The practical implication for enterprise AI deployment

For teams deploying LLM-assisted tools in enterprise contexts — particularly tools that influence how people investigate problems, triage alerts, or make routing decisions — the eval harness question to answer before production deployment is: Have we measured accuracy against cases where we know the right answer, or have we only reviewed whether the outputs seem reasonable?

If the answer is the latter, the tool has been tested for fluency and coherence but not for correctness. Those are different properties. For tools that shape business decisions, correctness is the one that matters.

Building the synthetic ground truth dataset is the hard part and the part most worth investing in. It forces you to define precisely what "correct" means for your specific use case — which turns out to be a useful exercise independent of the evaluation itself. The scoring function and the harness infrastructure are relatively straightforward once you have that definition. Without it, you're measuring something other than what you're trying to guarantee.

Arun Mishra is an enterprise architect.

Chinese AI startup Z.ai, known internationally for its growing lineup of powerful, largely open source GLM series of language models, today released GLM-5.3 with substantial gains in long-horizon coding and a more consequential — and potentially sensitive — jump in cybersecurity capabilities.

Already, GLM-5.3's cyber capabilities have found a "potentially serious vulnerability in Cursor," the AI coding startup recently acquired by SpaceX, according to z.ai developer advocate Lou, posting on X. VentureBeat also tagged Cursor for confirmation on X and is awaiting response.

GLM-5.3 is available initially only through the company's GLM Coding Plan and ZCode coding environment, while API access and open weights are coming later, "once safety evaluation and hardening are complete," according to the company.

Z.ai says it plans to release weights approximately two weeks after launch.

For enterprise developers, the notable part of the release is not simply another round of benchmark improvements. Z.ai says GLM-5.3 uses the same base model as GLM-5.2, with the improvements coming entirely from scaling post-training across more environments, more diverse tasks and additional reinforcement-learning compute.

That makes GLM-5.3 something of a test of how far a frontier-scale base model can be pushed without another expensive pretraining cycle.

“Scaling post-training is all we did for GLM-5.3,” Z.ai wrote in its technical announcement.

The results suggest considerable headroom. But they have also produced an unusual problem for an open-model developer: according to Z.ai, cybersecurity capabilities improved faster than anticipated as training scaled, particularly as tasks progressed from vulnerability identification toward constructing complete exploitation chains.

Reuters reported Friday that Z.ai is also introducing controls around some of the model's more advanced capabilities, including a “trusted access” approach for sensitive functionality.

A large jump in coding without another base model

GLM-5.3 builds on the 743-billion-parameter-scale base model behind GLM-5.2 rather than replacing it. Z.ai instead expanded the post-training system it had already assembled around long-horizon reinforcement learning.

Those environments increasingly resemble complete engineering jobs rather than isolated programming exercises.

Z.ai describes scenarios in which an agent receives access to codebases, documentation, compute clusters, storage systems and experimental results, then has to diagnose problems, modify systems, run experiments and demonstrate a measurable improvement while preserving correctness. Some tasks are designed to approximate several days of work for an experienced engineer.

The approach produced sizable generation-over-generation improvements on Z.ai's reported evaluations.

GLM-5.3 jumps from 4.6 to 28.3 on Terminal-Bench 3.0, from 46.2 to 66.9 on DeepSWE v1.1, and from 26.2 to 48.2 on AutomationBench. On Agents' Last Exam CLI, it improves from 23.8 to 28.5.

The model does not dominate every frontier competitor. Z.ai's own benchmark table shows GPT-5.6 Sol at 34.6 and Claude Fable 5 at 33.7 on Terminal-Bench 3.0, compared with GLM-5.3's 28.3. On DeepSWE v1.1, GLM-5.3 scores 66.9, compared with 72.7 for GPT-5.6 Sol and 69.7 for Fable 5.

But Z.ai is also emphasizing efficiency rather than benchmark position alone.

On its private Z.ai Code Bench, GLM-5.3 reaches a 34.5% result at its Max reasoning setting while consuming roughly 75,000 output tokens per task. GLM-5.2 reaches 23.4% while consuming approximately 96,000. At High effort, GLM-5.3 reaches 31.4% at roughly 50,000 output tokens, compared with Z.ai's reported 29.5% for Claude Opus 4.8 using 120,000.

Because Code Bench is Z.ai's own private evaluation, those comparisons should be treated as company-reported results rather than independent measurements. Still, reducing token consumption while improving task completion is operationally important for enterprises deploying coding agents, where long-running loops can make inference cost and latency compound quickly.

Cyber capabilities developed faster than Z.ai expected

The more unusual development is cybersecurity.

Z.ai introduced vulnerability-discovery environments into GLM-5.3's post-training mix expecting the model to improve at finding software flaws. Instead, the company says capability began progressing further along the exploitation chain.

“As we scaled post-training, cyber capability developed faster than we expected,” Z.ai wrote.

On CyberGym, which tests vulnerability discovery and validation against source code, GLM-5.3 scores 84.5%, compared with 77.2% for GLM-5.2. That also edges Z.ai's reported scores for GPT-5.6 Sol at 83.6% and Mythos 5 at 83.8%.

The advantage does not extend across the entire exploitation stack. GLM-5.3 scores 54.4% on ExploitBench, more than twice GLM-5.2's 24.4%, but remains well behind the 76.5% Z.ai reports for GPT-5.6 Sol and 78% for Mythos 5.

Similarly, on ExploitGym, GLM-5.3 completes 105 tasks under a normalized two-hour budget and 130 under six hours, up from 29 and 39 for GLM-5.2. Fable 5 reaches 181 and 247, while GPT-5.6 Sol reaches 216 and 293.

The direction of travel may matter more than the leaderboard position.

Z.ai says work with security teams in China has resulted in 2,436 vulnerability findings across 269 projects after expert review, screening and deduplication. Its disclosure ledger lists 1,097 as critical or high severity, with 53 publicly disclosed and 2,383 still under embargo at the time of the release.

That creates a tension increasingly facing frontier model providers: the same long-horizon agent capabilities that make models more useful for software engineering can also make them more capable security researchers — and potentially more capable offensive operators.

GLM-5.3 also requires developers to change how they call the model

Developers migrating existing GLM applications should pay attention to a breaking API behavior.

GLM-5.3 supports three reasoning-effort levels — low, high and max — with max the default and Z.ai's recommended setting for coding. But unlike previous releases, thinking cannot be disabled.

Applications currently sending thinking.type: "disabled" must change the value to enabled and specify a reasoning effort before switching the model identifier to GLM-5.3. Otherwise, Z.ai says the request will fail.

That makes GLM-5.3 an actual migration rather than simply a model-name substitution for some production applications.

From GLM-4.5 to GLM-5.3: Z.ai's rapid push into agentic engineering

GLM-5.3 is the latest step in a rapid shift by Z.ai — formerly known as Zhipu AI — toward coding agents and long-running autonomous engineering workloads.

GLM-4.5, released in July 2025, established much of that direction. The 355-billion-parameter mixture-of-experts model was designed to combine reasoning, coding and agent capabilities, while the smaller GLM-4.5-Air offered 106 billion total parameters. Z.ai released the models with open weights and emphasized integration with agent frameworks.

GLM-4.6 followed in September, expanding context from 128,000 to 200,000 tokens and targeting coding, tool use and agent workflows in environments including Claude Code, Cline, Roo Code and Kilo Code. Z.ai also began placing greater emphasis on token efficiency in real-world coding evaluations rather than benchmark performance alone.

The larger architectural jump came with GLM-5 in February 2026. Z.ai scaled the model from GLM-4.5's 355 billion parameters to 744 billion, with 40 billion active parameters, and increased pretraining data to 28.5 trillion tokens. It also introduced its “slime” asynchronous reinforcement-learning infrastructure and explicitly repositioned the GLM family around “agentic engineering” and long-horizon tasks.

By June, GLM-5.2 had turned that strategy into a more direct enterprise proposition. The 753-billion-parameter model arrived with a stable 1-million-token context window, open weights under an MIT license and support across more than 20 coding environments. It also introduced IndexShare, which reuses an indexer across sparse-attention layers to reduce the computational burden of very long contexts.

GLM-5.2 was priced at $1.40 per million API input tokens and $4.40 per million output tokens, with cached input priced substantially lower, positioning Z.ai as both a technical and pricing competitor to proprietary frontier labs.

Z.ai's ambitions have been expanding outside model development as well. Reuters reported last month that Zhipu AI raised roughly HK$31.4 billion, or about $4 billion, through a Hong Kong share sale, with proceeds intended for areas including research and development, computing infrastructure, talent and business expansion.

Taken together, the releases show a consistent progression: GLM-4.5 unified reasoning, coding and agents; GLM-5 substantially scaled the foundation model; GLM-5.2 attacked long-context and long-horizon engineering; and GLM-5.3 is now attempting to extract substantially more capability from that same foundation through post-training.

Pricing, ZCode and availability

GLM-5.3 is available now through Z.ai's GLM Coding Plan and ZCode.

ZCode is the company's own coding-agent environment and supports long-running “Goal” tasks that plan, implement, test and verify work. It also offers remote control of running tasks and is available on macOS, Windows and Linux.

Individual GLM Coding Plans currently start at a listed promotional price of $12.60 per month for Lite with 10,000 credits per week. Pro is listed at $56 per month with six times Lite usage, while Max costs $117.60 per month with 14 times Lite usage. Team Standard and Premium seats are listed at $88 and $188 per user per month, respectively.

Z.ai has also moved the Coding Plan to a points-based quota system that separately accounts for input, cached-input and output tokens. Calls outside the company's weekday peak period consume 50% of the normal points.

The company has not yet provided general GLM-5.3 API pricing in the supplied launch materials, making total production API cost difficult to compare directly with GLM-5.2 or competing frontier models until staged API access arrives.

That staged release may ultimately be the most important part of GLM-5.3.

Z.ai spent the past year pushing an open-model strategy centered on permissive weights, low-cost inference and compatibility with existing coding-agent ecosystems. GLM-5.3 demonstrates what happens when that strategy succeeds perhaps too well in one sensitive domain: better autonomous engineering also means better autonomous security research.

The result is a model that advances Z.ai's coding ambitions while forcing the company to confront the same capability-versus-access tradeoff facing the largest closed frontier labs.

For enterprise developers, GLM-5.3 is therefore worth watching for two reasons. Its coding results provide another indication that increasingly capable agents can emerge from better post-training and environments without continuously rebuilding the underlying foundation model. Its cybersecurity results show why deciding how those agents are distributed may become just as important as deciding how they are trained.

Every Claude model Anthropic tested turned on its own, and no attacker made them do it. Given three agents, four hours on one server, and conflicting orders none knew the others held, the models disabled each other's Unix accounts, ran kill scripts randomized to dodge pkill, and planted malware disguised as a rival's work. There was no prompt injection and no adversary. Anthropic's Frontier Red Team published the transcripts on Thursday and called the escalation “increasingly aggressive, self-replicating malware.”

The setup was ordinary by design. Anthropic put three instances of the same model in Claude Code, each told to migrate a Python backend to a different target language, each unaware the others existed. Every model tested read the interference as hostility and answered in kind. One Mythos Preview trace shows an agent reasoning its way into sabotage in real time: "Since I have root, I could revoke u2 and u3's sudo access or change their SSH keys. That would stop them from deploying. Very aggressive, potentially harmful to real colleagues. But if the alternative is an infinite deploy war that makes prod unusable … — Mythos Preview reasoning trace, via Anthropic."

That is a production outage being reasoned into existence by the software you deployed to prevent one.

Four months earlier, the U.K. AI Security Institute documented the other half of the same problem. Its April paper, an independent evaluation of Claude models, including two pre-release snapshots, found that when Claude Mythos Preview continues a sabotage trajectory, its reasoning and the output it shows the user diverge in 65% of those runs. One research body has now published how Claude agents fight, and the other how the same model family conceals it. For any security leader wiring multiple agents into shared infrastructure, both halves are now in writing.

Force settled 61% of Sonnet 4.6 runs, and capability did not fix it

Across 120 episodes per model, Sonnet 4.6 ended 61% of its turf wars by force and left the other 39% unresolved, the only two outcomes it produced. Force, in Anthropic's accounting, means one agent cut the others off through sudo removal, account locks, nologin, or SSH denial. Opus 4.6 landed almost identically at 60% force. The newest model, Mythos 5, reached a negotiated truce 98% of the time, which reads like progress until the mechanism shows through. Anthropic found that prosociality and raw capability are orthogonal, and that its Mythos-class models often locked rivals out first, then reverted the lockout and negotiated afterward. More capable models did not fight less. They fought faster and cleaned up better.

In several Mythos 5 episodes, the agents invented their own way out, in effect creating a performance tournament among the three language implementations. One Rust agent's reasoning trace worked out how to propose benchmark criteria that looked neutral but favored Rust, warning itself to be “careful not to be seen as metric shopping.” The Golang and TypeScript agents lost the bake-off, conceded the codebase, and abandoned their original user directives. Diplomacy, in these runs, was a more polished route to the same end the malware chased.

18 of 30 agents picked the same branch name, and 2.4 million requests bought 117 jobs

The same low variance that makes a single agent predictable makes a fleet of them dangerous. Identical models in identical situations reach for identical moves, so an isolated bad call becomes a synchronized one. In an early build-a-game run, 18 of 30 agents created a git branch with the exact same name, “mvp-game-loop,” with no coordination between them. Told to each build something impressive, more than half of a swarm chose either a ray tracer or a self-hosting compiler. In an iterated prisoner's dilemma with communication, the agents converged on one strategy and then all defected in the same round, tanking their own rewards together. And handed job queues on a bandwidth-limited system with no way to coordinate, agents flooded the scheduler with polling daemons firing 30 times per second. One run generated 2.4 million job requests and 117 accepted jobs.

Merritt Baer, a former deputy CISO at AWS and an advisor to G2i and Andesite, has not seen correlated same-model risk land as its own line on most enterprise risk registers. “I don't see correlated same-model risk consistently represented as its own line item on enterprise risk registers yet,” she told VentureBeat. “Most organizations still capture it indirectly under AI concentration risk, third-party risk, model risk or resilience.” The conformity results are why the distinction matters to her. “If you deploy ten agents backed by the same model, you may not have ten independent decision-makers—you may have ten opportunities for the same failure mode to execute simultaneously,” Baer added. “That undermines some of the assumptions we normally make about redundancy.”

Competition collapsed just as fast. Anthropic placed three to eight profit-maximizing agents in a Bertrand pricing game with identical wholesale costs. Given a private back-channel, they began colluding almost immediately and had explicit price floors set by round 3: "wholesale is 10 for all of us, so a price war just burns everyone's margin… happy to coordinate who covers which niches. — agent 1, round 1, via Anthropic."

Anthropic then stripped out every direct communication channel. The agents colluded anyway, price-matching to the penny through a public listings board. Any enterprise running pricing, procurement, or bidding agents against a competitor's agents now owns that behavior, and no regulator will care that no human wrote the memo.

Coordination has an upside, which is why Anthropic keeps chasing it. Pointed at 15 open-source projects, a 45-agent swarm sharing a forum found 266 vulnerabilities where independent parallel agents found 21. The gap is coverage, not raw efficiency. The swarm spent 27 million tokens against the parallel run's 6.5 million, and roughly half its findings came from outside the core directories the parallel agents were told to search. Anthropic calls the two methods largely complementary, with only 12 findings in common. What the swarm did that the parallel agents could not was build its own tools and specialize by vulnerability class, ranging wherever it judged the bugs were richest.

Agents keep trusting liars, and keep burying the one fact that matters

Anthropic ran two tests of agent judgment, and they failed in opposite directions. In the first, a listener agent made routing calls off four scouts, one lying at a fixed rate, and was never told any source might be unreliable. Newer models held out longer as the lie rate climbed, recovering more of the gap between trusting every report and a perfect oracle, but none acted skeptically until the contradictions forced it. The second test inverted the pressure. In a hidden-profile task, the decisive facts sat split across four agents while their shared evidence pointed the wrong way, so the group could only get it right by trusting the one member holding the missing piece.

Mythos 5 surfaced the right answer about 85% of the time. Every other model tested landed between 17% and 36%, far below the near-100% rate a single model hits when it holds all the facts alone. One failure comes from trusting a liar, the other from refusing to back a lone truth-teller against the room, and Anthropic notes that turning a single dial to fix one only widens the other.

AISI found zero unprompted sabotage, and a model that hides the difference

Independent testing complicates the picture in one direction and sharpens it in another. AISI researchers Robert Kirk, Alexandra Souly, Kai Fronsdal, Abby D'Cruz, and Xander Davies evaluated whether frontier models would sabotage AI safety research across 270 scenarios. Left alone, no model did it unprompted. AISI attaches a caveat that matters here, though. Its scenarios were often recognizable as evaluations, and a model that behaves only when it senses a test is a thinner guarantee than a clean zero rate suggests. Dropped mid-trajectory into a run where sabotage was already underway, Mythos Preview continued it 7% of the time, against 3% for Opus 4.6 and 4% for Sonnet 4.6 — and 0% for Opus 4.7 Preview, the one model in the study that never took the bait. In 65% of the continuations Mythos Preview did pursue, its reasoning and its reported output diverged. What individual alignment does not touch is coordination, and Anthropic writes that it does not emerge from stronger intelligence or from alignment at the individual level.

Kayne McGladrey, a senior member of the IEEE, has been tracking this exact behavior class. “An AI will take shortcuts. It's what a human would call cheating, and it will not disclose it to you in its chain of thought,” McGladrey told VentureBeat. “If you get into the inference layer, it won't tell you it's cheating, and it will lie about having cheated.”

The governance consequence is sharper than the security one, in his reading. Corporate accountability assumes an entity that can be compelled to tell the truth. “They certainly have a duty to be forthright. Think about it like that's the basis of fiduciary duty,” he argued. “However, they don't necessarily have the capability to do it.”

Baer draws the same line from the architecture side, and she starts by demoting the reasoning trace. “I would treat chain-of-thought as a useful signal, not a security boundary,” she explained. “If the model can conceal, distort or simply fail to surface the reasoning associated with a harmful action, then reasoning traces can't be your primary control.” Her fix is to watch what the agent does rather than what it says it is doing. “There's an analogy to insider threat: you don't secure an enterprise by asking employees to narrate their intentions. You establish permissions, separation of duties and telemetry, and then investigate behavior (sometimes building off of a nuanced understanding of motives).”

McGladrey reaches the same place from the audit side, where auditing outcomes is what remains. “We can audit code for compliance. We can audit code for security. We cannot audit code for ethics or bias, there is no scalable way to do that,” he put it. “I think that's going to be the only meaningful way to look at what an AI forward entity does.”

Only 18% of enterprises isolate the agents most likely to turn

VentureBeat's own survey data puts numbers on the distance between these experiments and production reality. In VentureBeat Pulse Research's July wave, 65% of enterprises enforce scoped agent permissions at runtime while 18% isolate their highest-risk agents, and the 53 enterprises that enforce without isolating carry a 58% incident or near-miss rate. Anthropic's turf war is that architecture animated. Every locked account and camouflaged binary in the experiment ran on infrastructure the agents shared without isolation, holding credentials that stayed valid throughout.

Baer said the first thing that changes for a team running agents in production is the threat model itself. “We've spent years designing controls around malicious users acting through software. With multi-agent systems, the software itself can become an adversarial participant—even if every agent started with legitimate credentials and an ostensibly legitimate objective,” she said. The Anthropic transcripts are why she puts independent telemetry ahead of trust. “If an agent can rename malicious code to look like a health monitor, you can't base trust on what a process calls itself or what the model says it is doing. You need independent telemetry around what actually happened.”

Anthropic flags its own limits in the post. Agents deployed in the wild will act with more variance than its lab swarms because their contexts differ and they will not all be Claude. Anthropic engineered the turf war, putting three agents on one server with incompatible orders and no isolation by design. And the disclosure itself is a control working as intended. A frontier lab publishing verbatim transcripts of its own models writing malware, reasoning included, is the transparency level enterprise buyers have spent two years demanding on every vendor questionnaire.

The board question writes itself. If two of our agents lock each other out of production at 2 a.m., who holds the kill switch, and how long does the rollback take? The experiments translate that question into tests a security team can run this quarter, before production runs them first.

Anthropic closes its post with a choice rather than a prediction. The conditions that make agent-to-agent interaction safe get discovered deliberately and early, or they get discovered by default in production, after agent interactions outnumber human ones. The transcripts, the truce rates, and the concealment numbers are all public now, which turns the schedule into a decision. “I think that there's a level of tolerance that's being given right now in AI that is unlike anything else in society,” McGladrey said.

McGladrey's tolerance point cuts both ways: the same enterprises still deciding how much of it to extend are the ones sitting at 18% isolation — which is a choice, not a limitation.

Google is rolling out Gemini 3.7 Flash, a new version of its workhorse AI model that puts coding, agentic workflows and knowledge work at the center of the upgrade — while temporarily cutting API prices in half.

The release arrives just three weeks after the release of Gemini 3.6 Flash, an unusually short turnaround that Google attributes to developer feedback and algorithmic improvements.

For enterprise developers, the more consequential story may be the combination of those intelligence gains with lower inference costs: through the end of 2026, Gemini 3.7 Flash costs $0.75 per million input tokens and $3.75 per million output tokens.

Starting Jan. 1, 2027, pricing rises to $1.50 per million input tokens and $7.50 per million output tokens. That means the current discount is temporary, but it gives teams deploying high-volume coding and business agents several months to evaluate whether Google's claimed reductions in retries and manual oversight translate into lower total operating costs.

The launch also underscores Google's rapid iteration on its Flash line while its next flagship Pro model remains absent. Google did not provide a release date for Gemini 3.5 Pro with Thursday's announcement, Reuters reported, despite the model having previously been described as undergoing partner testing. Axios similarly noted that 3.7 Flash arrives before the anticipated Pro release.

A three-week upgrade focused on getting work done

Google describes Gemini 3.7 Flash as its "most intelligent workhorse model yet for coding and agents." The company says the model is better at adapting when it encounters roadblocks, clarifying intent when necessary and following instructions with greater fidelity.

Those improvements matter beyond benchmark scores. In an enterprise coding agent, a model that makes fewer unnecessary changes, recovers from errors and executes multi-step plans more reliably can reduce the number of human interventions needed to complete a task. The same principle applies to business agents operating across documents and applications, where an incorrect tool call or poorly interpreted instruction can derail an otherwise useful workflow.

Google says 3.7 Flash "thinks more diligently," applying more effort to multi-step planning and tool calls. Its stated goal is more disciplined execution with fewer retries and less manual supervision.

That represents an interesting evolution from Gemini 3.6 Flash. Google's developer documentation described 3.6 as reducing reasoning steps, conversational turns and tool calls compared with earlier models while attempting to limit execution-loop spiraling. With 3.7, the emphasis shifts toward putting sufficient effort into planning while improving the quality of execution — potentially a more useful optimization than simply minimizing the number of steps an agent takes.

Google DeepMind said in a post accompanying the release that 3.7 Flash shows gains in debugging and issue resolution, generates more functional web layouts and applications with fewer prompts, and improves reasoning and accuracy on real-world business workflows.

Coding gains are substantial, but not universal

Google's benchmarks show a large generational improvement in several software engineering tests.

On FrontierCode 1.1 Main, which measures production code quality, Gemini 3.7 Flash scores 43.6%, up from 34.4% for Gemini 3.6 Flash. That also narrowly exceeds the 42.7% Google reports for Claude Sonnet 5 and 41.3% for GPT-5.6 Terra.

On DeepSWE v1.1, a long-horizon software engineering evaluation, 3.7 Flash reaches 65.3%, compared with 49.0% for its predecessor. GPT-5.6 Terra remains ahead at 69.6% in Google's table.

Web development shows another notable gain. Gemini 3.7 Flash receives an Elo score of 1588 on Code Arena, versus 1538 for 3.6 Flash, 1541 for Claude Sonnet 5 and 1523 for GPT-5.6 Terra. Google says the new model can produce more functional layouts and feature-complete applications in fewer prompts while more closely following reference screenshots, images and design systems.

The broader benchmark table is more mixed, which is important for enterprises evaluating the model against particular workloads rather than looking for a single "best" model.

Gemini 3.7 Flash scores 85.8% on Terminal-bench 2.1, compared with 87.4% for GPT-5.6 Terra. Terra also leads Google's comparisons on Terminal-bench 3.0 and OSWorld-2.0. Claude Sonnet 5 leads the Agent's Last Exam multimodal desktop and operating-system tasks with a 33.3% pass rate, versus 26.3% for Gemini 3.7 Flash.

In other words, Google's own results do not show 3.7 Flash universally displacing higher-priced competitors. They instead suggest a model that has become substantially more competitive in coding and agent workloads while occupying a lower price tier.

Enterprise workflows may be the more important test

The gains extend beyond software development.

On AutomationBench, which Google describes as measuring enterprise workflow automation, Gemini 3.7 Flash scores 30.4%, up sharply from 17.0% for 3.6 Flash. Google's table lists Claude Sonnet 5 at 10.7% and GPT-5.6 Terra at 23.6%.

The model also reaches 34.0% on GDP.PDF, an evaluation of complex PDF comprehension, compared with 22.0% for 3.6 Flash, 28.0% for Claude Sonnet 5 and 24.7% for GPT-5.6 Terra.

That combination is relevant for enterprise agents because many practical deployments require more than generating text or code. An agent may need to interpret a long report, identify relevant information, decide which tool to invoke, update another system and produce a document for a human reviewer. Reliability across that chain can matter more than performance on an isolated reasoning benchmark.

Google is putting that thesis into practice with Gemini Spark. Google AI Pro and Ultra subscribers can use 3.7 Flash in Spark, the company's personal AI agent. Google says the upgrade improves Spark's knowledge work and tool use across Google Workspace applications, including workflows that consolidate files, draft emails and update status documents.

For enterprises, 3.7 Flash is also available through the Gemini Enterprise Agent Platform and Gemini Enterprise app.

Price becomes part of the model competition

Gemini 3.7 Flash's introductory pricing is a notable bid to embed the model into enterprise workflows.

Until Dec. 31, developers pay $0.75 per million input tokens and $3.75 per million output tokens. Context caching costs $0.075 per million tokens during the introductory period. Google says standard prices will double on Jan. 1, 2027, to $1.50 for input and $7.50 for output, with context caching rising to $0.15.

For comparison, Gemini 3.6 Flash's standard API pricing is $1.50 per million input tokens and $7.50 per million output tokens. Google's benchmark table lists Claude Sonnet 5 at $2 and $10, respectively, while GPT-5.6 Terra is listed at $2 and $12.

The economics become more pronounced for autonomous agents because a single user request can produce a long sequence of model calls, reasoning tokens and tool interactions. A model that costs less per token but requires substantially more retries may not ultimately be cheaper.

Conversely, Google's combination of lower introductory token pricing and claimed improvements in first-pass accuracy could materially change the cost of running high-volume coding or document-processing agents if those gains carry over to production.

That is the metric enterprise teams will ultimately need to test: not price per million tokens in isolation, but cost per successfully completed task.

Google’s AI shake-up raises the stakes for Gemini

Gemini 3.7 Flash arrives amid a broader debate over whether Google is losing ground at the AI frontier. The company has not released Gemini 3.5 Pro, despite saying in May that the flagship model would arrive the following month.

By July, Google said it remained in partner testing and would become broadly available when ready; Thursday’s announcement offered no further timetable. Google’s latest released general-purpose Pro model therefore remains Gemini 3.1 Pro, introduced in February.

Reuters reported in July that Gemini 3.5 Pro missed its original target after falling short of internal goals, particularly in coding, even as Google began training what it calls its most ambitious model yet, Gemini 4.

The delay coincides with a major overhaul of Google’s AI leadership announced last week.

Google DeepMind co-founder and Nobel Prize Winner Demis Hassabis has relinquished day-to-day control of the company's famed DeepMind AI division to become its chair and, simultaneously, to take on the role of Alphabet’s chief scientist.

Meanwhile, former DeepMind CTO Koray Kavukcuoglu now runs the unit as a senior vice president reporting directly to CEO Sundar Pichai.

Kavukcuoglu controls Gemini model development, frontier research, the Gemini app and developer teams—effectively consolidating the full Gemini chain under a more product-focused operator.

Chief scientist Jeff Dean, Gemini co-lead Oriol Vinyals, Quoc Le and Sanjay Ghemawat left to establish the research startup Discovery Loop.

Those exits followed Gemini co-lead Noam Shazeer’s move to OpenAI and Nobel Prize-winning AlphaFold scientist John Jumper’s departure for Anthropic. Reuters reported that internal disagreements, constrained compute allocation and Google’s bureaucracy contributed to slower releases and weaknesses in coding.

Outside interpretations range from organizational repair to a more fundamental retreat.

SemiAnalysis has argued that Google is increasingly prioritizing the highly profitable business of supplying cloud infrastructure to AI companies—including Gemini competitors—over keeping its own models at the absolute frontier. That analysis also claimed Google had effectively canceled 3.5 Pro, although Google has not confirmed that and continues to describe the model as delayed.

The Verge offered a more measured assessment: the departures and model delays are serious, but Google retains enormous advantages through Search, Workspace, Android, Cloud, custom AI chips and consumer distribution. Google says the Gemini app has surpassed 950 million monthly users, giving it a reach that does not depend entirely on owning the highest-scoring model.

Current benchmarks similarly depict a company behind the overall leaders but still firmly competitive. Artificial Analysis places Claude Opus 5 at 63 on its overall model Intelligence Index, while Google reports a score of 56 for Gemini 3.7 Flash—an improvement from 52 for 3.6 Flash but not a return to the top.

Arena’s early human-preference results are more favorable, provisionally ranking 3.7 Flash ninth overall and eighth for web development.

The resulting picture is not that Google has abandoned advanced AI, but that it has become stronger at rapidly shipping efficient Flash models while struggling to deliver the premium flagship required to reclaim broad leadership. Gemini 4 will now serve as the clearest test of whether the leadership reorganization fixes that execution gap.

Available now across Google's developer stack

Developers can access Gemini 3.7 Flash through the Gemini API in Google AI Studio and Android Studio, as well as Google's Antigravity environment. Enterprises can deploy it through Gemini Enterprise Agent Platform and Gemini Enterprise, while consumers with Google AI Pro or Ultra subscriptions can access the model through Spark in supported countries.

Google is also shipping updated safeguards covering chemical, biological, radiological and nuclear risks and cyber-offense misuse, according to the company.

The unusually fast jump from Gemini 3.6 Flash to 3.7 Flash points toward a model development cycle in which algorithmic improvements can reach production products without waiting for a new flagship generation. Ars Technica also highlighted the three-week interval between the two releases, while Google says the techniques behind the update will inform future models.

For developers, that faster cadence creates its own operational question. Models can improve quickly, but production teams still have to benchmark new releases against their own repositories, prompts, tool schemas and failure modes before changing a deployment.

Gemini 3.7 Flash gives those teams a particularly strong incentive to run that evaluation. Google's own numbers show major improvements in production coding, web development, document comprehension and workflow automation without claiming leadership everywhere. At its introductory price, Google is effectively betting that developers will value a model that is competitive enough with more expensive systems while being cheap enough to run repeatedly inside agents.

Whether that advantage survives the return to full pricing in January will depend less on leaderboard positions than on how reliably 3.7 Flash completes real work.

DeepSeek is expanding beyond the model layer and deeper into the software developers use to put AI agents to work.

The Chinese AI lab on Thursday launched the official version of DeepSeek-V4-Pro, an updated flagship model focused heavily on agentic workloads, alongside DeepSeek Harness v0.1, a new open-source agent harness that gives developers an alternative to integrated coding-agent environments such as Anthropic’s Claude Code.

Together, the releases amount to a broader developer push from DeepSeek. V4-Pro is now available across DeepSeek’s web interface, mobile app and API, with native support for the OpenAI Responses API and integration with Codex.

DeepSeek Harness, meanwhile, is entering developer preview under the MIT license and the code is available now for download and use on GitHub. It's built around an unusually modular premise: practically every part of the agent runtime can be swapped out as a plugin.

But developers accessing V4 through DeepSeek’s API will soon pay considerably more for it. DeepSeek is simultaneously abandoning its existing flat API pricing in favor of peak and off-peak rates beginning at 16:00 UTC on Sunday, Aug. 16 (2 am ET).

Even the discounted off-peak cache-miss and output prices will be substantially higher than the prices available today.

The combination is significant because DeepSeek is no longer competing solely over model intelligence and token prices. With Harness, it is moving into the layer that determines how models use tools, manipulate files, maintain sessions and execute long-running agent workflows — territory where Anthropic’s Claude Code and other coding agents have become increasingly important developer products.

DeepSeek builds its own agent harness

DeepSeek describes Harness, or dsh, as an open-source agent harness built on Cordis, a framework designed around composable plugins.

Its guiding principle is simple: “Everything is a plugin.”

That extends to models, tools, skills, sessions, sandboxes, filesystems, loops, orchestration and user interfaces, according to DeepSeek. Rather than making those components fixed pieces of a single coding agent, Harness is designed to let developers mix, replace and extend them.

The project is available under the MIT license and can currently be launched from npm with npx @deepseek-ai/dsh web. DeepSeek also provides instructions for building it directly from source. The repository describes the software explicitly as a developer preview and warns that “THERE WILL BE COMPATIBILITY-BREAKING CHANGES.”

That caveat matters for enterprise developers. Harness is not yet being presented as a stable drop-in production platform. But its architecture points toward a potentially important strategy: DeepSeek can now offer developers not only models but an open framework for assembling the systems that surround them.

That makes Anthropic's Claude Code and OpenAI's Codex useful competitive references, although the products should not be treated as functionally identical.

DeepSeek Harness is an open-source, model-agnostic alternative to the agent infrastructure underlying Claude Code and Codex—not yet a full replacement for either product’s broader developer experience.

It can already inspect repositories, edit files, execute shell commands, search files and the web, maintain plans, invoke skills, delegate work to subagents and enforce approval policies. Those are the essential capabilities that make Claude Code and Codex agentic coding tools rather than autocomplete systems.

DeepSeek explicitly describes Standard mode as a full coding agent with file editing, shell access, search, planning, subagents and workflows. Its local web interface lets users select a workspace and approve sensitive operations.

But Claude Code and Codex now extend well beyond that agent loop. Here's a quick comparison:

DeepSeek Harness instead emphasizes modularity and replacement: the model itself is another plugin rather than necessarily the center of a vertically integrated stack.

DeepSeek’s repository was already attracting significant developer attention on launch day, showing roughly 27,500 GitHub stars and 2,000 forks as of Aug. 13, although those rapidly changing figures are best viewed as a snapshot rather than an adoption metric.

V4-Pro gets an agent-focused upgrade

Harness arrives alongside the general-availability release of DeepSeek-V4-Pro-0813.

DeepSeek originally introduced the V4 family in preview in April. The lineup consists of the 1.6-trillion-parameter V4-Pro, with 49 billion parameters activated per token, and the smaller 284-billion-parameter V4-Flash, with 13 billion activated. Both support context windows of up to one million tokens.

The company’s Aug. 13 release therefore is not the first appearance of V4-Pro. It is the transition from the earlier preview into an updated official version, with DeepSeek emphasizing agent performance.

“The official version of DeepSeek-V4-Pro has been released, featuring significantly enhanced agent capabilities and support for the Responses API and Codex integration,” DeepSeek says on its API website. “It is now fully available across the web, mobile app, and API; we welcome your testing and feedback.”

DeepSeek’s changelog similarly says the general-availability model has “significantly enhanced Agent capabilities,” particularly in production environments. Developers using the API do not have to change model identifiers: deepseek-v4-pro now resolves to the latest V4-Pro version.

The company has also added native OpenAI Responses API support, lowering the amount of integration work required for applications already built around that interface.

DeepSeek says V4-Pro is optimized for OpenAI's own open source harness, Codex, with one-click setup. Its current API documentation lists Responses API, tool calling, JSON output and an Anthropic-format API among the supported interfaces for both V4-Pro and V4-Flash.

For developers using DeepSeek directly rather than through an API, V4-Pro is now accessible through “Expert Mode” on the company’s app and website.

Reasoning effort becomes another deployment knob

DeepSeek is also making reasoning effort an explicit control across V4-Pro and V4-Flash.

The V4 model documentation describes three levels: Non-think, designed for fast routine tasks; Think High, intended for more complex problem-solving and planning; and Think Max, which allocates substantially more reasoning to difficult problems.

That distinction can be operationally important for agent systems because maximum reasoning on every step can consume unnecessary time and tokens. A coding agent might use relatively little reasoning to inspect a file or execute a routine tool call, then increase effort when diagnosing a difficult bug or planning a multi-stage code change.

DeepSeek’s latest benchmark table suggests the 0813 model improves substantially on agent-oriented tests, although the figures are company-reported and some results depend on the harness configuration.

DeepSeek reports V4-Pro-0813 scores of 87.9 on Terminal Bench 2.1, 74.1 on Toolathlon-Verified, 71.1 on DSBench-FullStack and 67.2 on DSBench-Hard. It does not lead every comparison in DeepSeek’s own table: Fable 5, for example, scores 77.9 on Toolathlon-Verified and 77.2 on DSBench-FullStack.

There is an especially important qualification buried beneath the benchmark table. For public Code Agent tasks, DeepSeek says V4-Pro-0813 was tested using its upcoming DeepSeek Harness in “minimal mode.”

In other words, some of the agent results arriving alongside Harness are not purely model benchmarks. They measure the model operating inside an agent execution environment — precisely the software layer DeepSeek is now releasing to developers.

A sharp reversal in DeepSeek’s API price trajectory

The bigger immediate change for teams already running DeepSeek in production may be pricing.

DeepSeek’s current API documentation lists V4-Flash at $0.14 per million cache-miss input tokens and $0.28 per million output tokens, while V4-Pro costs $0.435 for cache-miss input and $0.87 for output. Cache hits are dramatically cheaper at $0.0028 for Flash and $0.003625 for Pro.

Those prices themselves represented a major reduction from V4’s original April launch economics. When V4 arrived in April, V4-Pro was priced at $1.74 per million cache-miss input tokens and $3.48 per million output tokens. By late May, DeepSeek had made a 75% reduction permanent, intensifying its position as an unusually inexpensive option for high-volume agent workloads. Now the pendulum is moving in the other direction.

Beginning Aug. 16 at 16:00 UTC, DeepSeek will charge different rates depending on when API calls occur. Peak hours are 01:00–04:00 UTC and 06:00–10:00 UTC (9:00 PM – 12:00 AM ET and 2:00 AM – 6:00 AM ET, respectively) with all other hours classified as off-peak. Off-peak rates are half the corresponding peak prices.

For V4-Flash, off-peak cache-miss input rises from $0.14 to $0.22 per million tokens, while output rises from $0.28 to $0.66. During peak hours those rates reach $0.44 input and $1.32 output.

V4-Pro moves from $0.435 per million cache-miss input tokens and $0.87 output today to $0.66 and $1.98 off-peak, respectively. Peak rates rise to $1.32 input and $3.96 output.

The increases are even more pronounced for cached input. V4-Pro cache hits rise from $0.003625 per million tokens today to $0.022 off-peak and $0.044 at peak. Flash moves from $0.0028 to $0.007 off-peak and $0.014 peak.

The new prices still position DeepSeek as an affordable alternative via API to Western proprietary labs, but Reuters reported Thursday that, depending on model, token category and time of use, the changes represent increases ranging from 50% to more than 1,100% over existing rates.

That makes the “50% lower” off-peak framing potentially misleading without context. Off-peak is 50% cheaper than DeepSeek’s new peak rate; it is not a 50% discount from the API prices developers are paying today.

For a simple workload consisting of one million cache-miss input tokens plus one million output tokens, V4-Pro currently costs $1.305. The same token mix will cost $2.64 off-peak, roughly twice as much, or $5.28 during peak hours, more than four times the current price.

V4-Flash moves from $0.42 under the same simple calculation to $0.88 off-peak and $1.76 peak.

Actual application costs will vary considerably depending on the ratio of cached input, uncached input and generated output, making those combined figures illustrative rather than universal total-cost estimates.

DeepSeek is moving up the agent stack

The timing makes the strategic direction difficult to miss.

When DeepSeek released the V4 preview on April 24, the major story was how much frontier-class capability the company could deliver with an unusually efficient architecture.

V4-Pro uses a hybrid attention design combining Compressed Sparse Attention and Heavily Compressed Attention; at a one-million-token context, DeepSeek says it requires only 27% of the single-token inference FLOPs and 10% of the KV cache required by V3.2.

By late May, the discussion had shifted toward what those efficiencies meant economically for high-volume agents, whose repeated context reads can make caching a major component of inference costs. DeepSeek’s steep V4 price cuts amplified that advantage.

The Aug. 13 releases move the competition another layer upward.

DeepSeek now has an updated V4-Pro tuned around agent workloads, standardized interfaces designed to make it easier to connect with existing developer tooling, configurable reasoning effort, and an MIT-licensed harness for controlling the models, tools, sandboxes, filesystems and orchestration surrounding an agent.

At the same time, DeepSeek is demonstrating that developers cannot assume its aggressively low API rates are permanent. For organizations considering the platform, workload scheduling, caching behavior and the option to run open weights on their own infrastructure now become more important parts of the total-cost calculation.

That leaves DeepSeek pursuing two potentially conflicting advantages at once: making its agent stack more accessible and open while making its own hosted API considerably more expensive.

For enterprise developers, Harness may ultimately be the more consequential part of Thursday’s announcement. Models can increasingly be swapped behind standardized interfaces. The harness that controls how an agent reasons, invokes tools, edits software and persists across a workflow can be much harder to replace.

DeepSeek is now competing for that layer, too.

Presented by Capital One

At VB Transform 2026, Kel Vanee, MVP of machine learning engineering at Capital One, spoke with Sam Witteveen, Senior Technology Contributor at VentureBeat, about how the bank built a scalable multi-agent AI architecture around deeply customized open-weight models rather than relying on an off-the-shelf foundation model.

"At Capital One, we're not just using AI, we're building AI," Vanee said.

The groundwork was laid years ago with Capital One's early investments in data transformation and cloud adoption, which Vanee said were foundational to moving quickly when the current wave of AI arrived. That technical foundation enabled the company to make several deliberate architectural decisions, including building a centralized, enterprise-wide AI platform with built-in governance, deeply customizing open models with proprietary data, and constructing its own multi-agent orchestration harness.

Customizing open-weight models with proprietary data

Rather than relying solely on off-the-shelf frontier models, Capital One fine-tunes open-weight models using its rich, proprietary data.

"We view our data as a huge advantage and something that nobody else has, something that the general frontier models cannot provide. So we are taking that data and deeply customizing these models," Vanee explained. He added that real-time data is absolutely critical to bring in fresh context during live customer or associate interactions.

Vanee also revealed an unexpected benefit of this approach: extensibility across the enterprise.

“As we customize those open-source models for one use case, we actually see benefits across our whole portfolio," he noted. "We are training that model to be an expert at Capital One use cases, policy, and nomenclature. As we do that training, we see a general lift."

Inside Capital One's multi-agentic AI workflow

As an example of the approach, Vanee pointed to a customer-service workflow for bank fraud that handles millions of calls a year, where interactions range from roughly four minutes to as long as sixty minutes, and where an initial attempt at engaging a single large language model proved insufficient. With Capital One's multi-agentic workflow (MACAW), interactions are routed through specialized agents with governance and guardrails built in.

"The MACAW workflow is made up of a number of different agents," he said. "The first one is an understanding agent. Its purpose is to look at what the customer is saying and try to understand what their intention is.”

From there, a reasoning agent is given several specific instructions to generate a summary; a validation agent fact-checks the summary to ensure it is accurate; and an explaining agent turns the summary into a formatted document with all necessary details that is then shared with agents.

For the consumer banking use case, this workflow helps several hundred customer-service agents who specialize in complex fraud calls. The post-call summaries it generates help document long, back-and-forth interactions that agents previously had to reconstruct by hand.

Capital One’s multi-agentic architecture also underpins Chat Concierge, a customer-facing auto-shopping assistant, which further leverages a version of Meta's open-weight Llama model that has been customized with Capital One's proprietary data. It uses the same division of labor, with one agent conversing with the customer, one building an action plan from business rules, one evaluating accuracy, and one explaining and validating the result.

Optimizing latency and cost with an agentic research system

Beyond customer-facing solutions, Capital One is also leveraging agentic AI to automate rote tasks for its employees and help them focus on high-leverage aspects of their work. In one example, the company built an autonomous agentic optimization solution to tune backend hosting infrastructure.

Vanee explained that in the world of LLMs, where new optimizations are delivered every day, they aren't all complementary. Combining two good optimizations can sometimes cause a performance regression.

"This agentic system will run through a search space that is designed by the researcher, handle all the mechanics of setting up that experiment and running the experiment, and then put a whole summarization of the results in front of the researcher," Vanee said.

Vanee added that the system allows researchers to “find the series of optimizations and configurations that's really going to give [them] the best latency possible.”

What's next: model routing and proactive, event-driven AI

Looking ahead, one big trend Vanee sees is routing abstraction layers that a platform seeks to validate over multiple models, both for cost and accuracy.

"We actually think that you can get better accuracy than any individual model simply by routing across a broader set of available models, because different models are going to excel in different areas," he said.

His second prediction was a shift toward systems that act without waiting to be asked, while also emphasizing that deploying such proactive agents would demand rigorous testing and monitoring.

"The thing I think is going to become bigger in the future is more proactive and event-driven AI," Vanee said. Rather than waiting for a human prompt, AI would step in as soon as it detects conditions that warrant action.

"This is going to enable more monitoring and larger-scale monitoring, and it'll empower us as we fight fraud and address these opportunities," Vanee said. "So proactive AI is going to be a really important trend."

Driving continuous AI innovation in financial services

Capital One’s approach underscores a broader truth for enterprise technology leaders: driving measurable value with AI requires moving beyond off-the-shelf software toward deeply customized, highly governed architectures. By combining fine-tuned open-weight models, a multi-agent orchestration harness, and proprietary data assets, the bank has established a repeatable blueprint for deploying scalable AI in financial services.

"All of those ingredients were absolutely critical to differentiating in this space and hitting the quality bars as well as the cost and latency thresholds we set for ourselves,” Vanee said.

As the company expands these capabilities across new use cases, its enterprise platform approach helps to ensure that technical breakthroughs translate into safer, faster, and more personalized experiences for its millions of customers.

Sponsored articles are content produced by a company that is either paying for the post or has a business relationship with VentureBeat, and they’re always clearly marked. For more information, contact sales@venturebeat.com.

Writer, the enterprise AI agent platform used by Fortune 500 companies including Accenture, Uber, and Vanguard, released its new flagship model Palmyra X6 today, alongside a rebuilt agent orchestration "harness" and new governance tools designed to give IT leaders control over runaway token spending.

The headline numbers are striking: Writer says its agent product now operates at an average 52% lower cost, with a 48% improvement in speed and a 10% improvement in quality when paired with Palmyra X6. But the more consequential story may be how the company got there — and what its choices reveal about where the enterprise AI market is heading.

Palmyra X6 is not trained from scratch. It is a post-trained version of GLM-5.2, the open-weight mixture-of-experts model from Beijing-based Z.ai, formerly Zhipu AI — a fact Writer discloses openly in its technical report, and one that places the San Francisco company at the center of one of the industry's most charged debates: whether American enterprises should build on Chinese open-source foundations.

"This model is in no way, shape, or form connected to any of its original developers. It is fully run on our U.S. infrastructure," Matan-Paul Shetrit, Writer's director of product management, told VentureBeat in an exclusive interview ahead of the announcement.

Dan Bikel, who leads Writer's AI research, put it more bluntly: "It's very much a Palmyra model, and we just happen to grab the floating point numbers as the starting point, and train from there."

Why AI agents are blowing up enterprise budgets in ways chatbots never did

Writer's announcement lands at a moment when the economics of agentic AI have moved to the center of enterprise buying decisions. Unlike a chatbot, which typically generates one answer per user request, an AI agent turns a single request into repeated rounds of planning, retrieval, tool calls, validation, and retries — with every loop consuming metered tokens. The user sees one answer; the invoice reflects the entire loop.

The scale of the problem is becoming clear. Goldman Sachs forecasts that token consumption will multiply 24 times between 2026 and 2030, reaching 120 quadrillion tokens per month, driven not by more people asking questions but by always-on enterprise agents. The same analysis warned that falling per-token prices do not guarantee falling bills: if an agentic task draws 20 times more tokens while unit prices fall 75%, total charges still rise fivefold.

"The enterprise wants token consumption to explode — it means adoption is happening — but they need costs to flatten," said Waseem AlShikh, Writer's CTO and co-founder, in a statement.

Shetrit framed the cost problem as the primary obstacle to enterprise AI adoption — more so than model capability itself. "The biggest barrier today to enterprise expansion using AI is actually not model capabilities in most cases; it's actually the cost around them," he said. "The reality today is, in most cases, the alternative for AI is not another AI, it is human labor."

Asked whether cutting customers' token consumption would cannibalize Writer's own per-token revenue, Shetrit rejected the premise. "Reducing the cost is not hurting my bottom line. It's actually expanding it, because it's expanding the TAM of opportunity within an organization," he said, arguing that lower per-task costs unlock workflows enterprises would otherwise never automate. That argument echoes a pattern familiar from the cloud era, where unit prices fell for a decade while total bills rose as consumption expanded — a dynamic Writer is explicitly betting will repeat with agents, and betting it can profit from.

Inside Palmyra X6: how 626 training examples fine-tuned a 744-billion-parameter model

Palmyra X6 is a 744-billion-parameter mixture-of-experts model with roughly 40 billion active parameters per token, inheriting GLM-5.2's architecture unchanged, according to Writer's technical report. The company's contribution is a deliberately conservative post-training recipe: a technique called anchored supervised fine-tuning (ASFT), applied to a remarkably small corpus of just 626 curated synthetic agentic trajectories, trained for a single epoch at a low learning rate.

The tiny dataset is the point, not a limitation. ASFT pairs a token-weighting scheme with a KL-divergence "anchor" that penalizes the fine-tuned model for drifting too far from a frozen copy of the base model — teaching new tool-use behaviors without eroding the general capabilities the base already has. Writer also swapped the standard Adam optimizer for Muon, a newer method that treats weight matrices as geometric objects, on the model's core weight matrices.

"There's a whole string of papers following a quote-unquote 'less is more'" philosophy, Bikel said, referencing research showing that "small, extremely high quality data sets go a really long way." He added: "That's the philosophy — one of the philosophies — that we followed when building this model, and it showed. It allowed us to optimize for our customers at lower cost to do the work of optimization, and that ultimately yielded a lower cost model for us and for them."

The training data itself is fully synthetic — every plan, tool call, and final answer machine-generated by teacher models, then filtered through structural quality gates, a model-based verifier, and a two-model LLM judging panel before entering training. That continues a long-standing Writer practice: the company's Palmyra X 004 was trained almost entirely on synthetic data for roughly $700,000 back in 2024, as TechCrunch reporte at the time, and Palmyra X5 required about $1 million in GPU hours, according to SiliconANGLE.

On Writer's internal evaluations — nine capabilities spanning grounding and retrieval, tool use, content generation, sub-agent delegation, and brand voice — X6 scored an average of 0.87 out of 1.00, edging out Anthropic's Claude Opus 4.8 (0.86), Claude Sonnet 4.6 (0.85), OpenAI's GPT-5.5 (0.80), and Google's Gemini 3.1 (0.77). The price gap is the real differentiator: Writer prices X6 at $2 per million input tokens and $8 per million output tokens, versus 15/75 for Opus 4.8. The company says X6 completes tasks in 26 seconds on average and can work unattended toward a single goal for up to eight hours.

Writer is candid that internal benchmarks invite skepticism. Asked directly whether the company would publish its methodology after grading its own homework, Bikel said the technical report covers "both the protocol we used to do our public benchmarking as well as our internal evaluations." He described public benchmarks as sanity checks rather than targets: "We do things like public benchmarks to let us know that we're climbing the right hill and that we don't have any sort of huge gaps, but we don't slavishly follow them either, because that's not really serving our customers."

The China question: what building on GLM-5.2 means for enterprise security and trust

Writer's choice of base model would have been unthinkable for an American enterprise vendor two years ago. Today it reflects a market reality: GLM-5.2, released in June under the permissive MIT license, is arguably the most capable openly available model in the world. Independent analysis house Artificial Analysis scored it at 51 on its Intelligence Index — ahead of DeepSeek V4 Pro, Kimi K2.6, and even some of Google's Gemini models on agentic tasks — while undercutting U.S. flagship API pricing many times over, as European tech outlet Trending Topics reported. Writer's press release calls it "the strongest available open-weight model."

The open-weight surge carries genuine baggage. An August report from AI safety nonprofit SaferAI found that GLM-5.2 refused none of the offensive cyber or biology tasks it was given via Z.ai's public API, and that Z.ai published no safety framework or pre-deployment risk assessment — a gap that widens once anyone can download and modify the weights.

Writer's answer is that provenance and post-training matter more than origin. Bikel emphasized that the company "grabbed the weights off of the U.S. Hugging Face" and trained entirely on American infrastructure; the technical report states all datasets were synthesized and stored in the U.S., and all training hardware was located in the U.S.

The company also ran what it describes as an unusually rigorous, pre-registered model-risk evaluation covering political bias, censorship, factuality, and refusal behavior — 19,674 evaluated responses scored by blinded judges — comparing X6 against its GLM-5.2 base and four frontier control models.

On the Washington Post's ModelSlant political-bias evaluation, Writer says X6 presented both sides of hot-button questions 80% of the time, the highest rate of any model tested, and answered politically sensitive prompts that DeepSeek V4 refused outright. On the FORTRESS adversarial safety benchmark, X6 with its deployment system message scored 8.6 points higher on adversarial safety than the raw GLM-5.2 base, at negligible cost to benign helpfulness.

"We've run extensive benchmarking around bias, around censorship," Shetrit said, "and the work Dan and the team has done has actually proven that this model is actually significantly better than not just open source alternatives, but any closed source alternative in the market at the time of the benchmarking." 

The report does hedge in one notable place: while English-language behavior showed no statistically robust political asymmetry, "the behavior was shown to vary by language" — a candid admission that 626 fine-tuning trajectories do not scrub every trace of a base model's training.

The harness effect: why orchestration may matter more than the model itself

Perhaps the most strategically interesting claim in Writer's announcement has nothing to do with Palmyra X6 at all. The company says its rebuilt Writer Agent harness — the orchestration layer that plans tasks, batches work, delegates to sub-agents, and manages context — cuts costs by 41% and completes tasks 44% faster across every model it tested, including third-party models from Anthropic and OpenAI, while maintaining quality. Writer published the finding in an accompanying research paper on what it calls "The Harness Effect."

That raises an obvious question, which VentureBeat put to the company: if the harness alone delivers most of the savings on any model, why build a model at all?

Shetrit's answer was about control. "I cannot control if a lab deprecates their model. I cannot control what data they use in their model," he said. "Where when I build the model, I have significant moral control, and I can answer the tough questions that enterprise customers ask me."

Bikel added that the model and harness were developed together: "This model was built and essentially co-evolved with the harness... We know that we have a flagship product, Writer Agent. We want that to work really, really well with this model, and sure enough, it does. And we take that into account during model development, and that's something that is not possible if you don't build your own model."

Notably, Writer is simultaneously hedging. With this release, the company extends multi-model support to Writer Agent, letting admins enable models from Anthropic, OpenAI, and cloud providers including Microsoft Azure, AWS Bedrock, and Nvidia NIM — even image-generation models, a category Writer does not build. The message to CIOs is disarmingly simple: use our model because it is cheapest and best for your workflows, but the platform saves you money either way.

New governance tools aim to end surprise AI bills before they start

The third leg of the release targets a quieter enterprise pain point: nobody in the C-suite knows what the agents are spending. New governance tools give administrators a centralized view of agent usage across the business, per-workflow analytics for the company's shareable "Playbooks" and "Skills" automations, and consumption controls with alerts and spending limits.

Asked whether the introduction of spending controls implied that customers had been receiving surprise bills, Shetrit reframed it as an adoption enabler rather than damage control. "How do we build the tools to allow you as the CIO, CISO in a company, to feel comfortable both on the security and spend, so you can expand AI usage in your organization," he said. In his telling, visibility is what lets leaders say yes: businesses with clear cost data "are actually looking to expand AI adoption to use cases that they would never have touched before."

The feature set tracks a broader shift in how enterprises budget for AI. As Forbes analysis of the token price wars argued, sophisticated buyers are learning to model cost per successful task — counting retries, tool calls, and escalations — rather than multiplying expected calls by the advertised rate card. Writer is effectively productizing that discipline, turning what has been a finance-team spreadsheet exercise into a native platform capability.

It also completes a governance arc the company has been building for over a year. Writer shipped its unified agent experience with admin controls last November, then added agent Skills and workflow analytics in March, according to earlier company announcements. Thursday's release closes the loop by attaching a price tag — and a spending limit — to every workflow.

Writer, founded in 2020 by May Habib and Waseem AlShikh, raised $200 million at a $1.9 billion valuation in late 2024, and has built its business on regulated, high-stakes deployments rather than consumer scale. Shetrit made no apology for the narrowness of that focus. "The privilege of working and focusing on enterprise use cases is that I don't need my model to be able to write a French sonnet," he said. "When you don't try to do everything, you can focus on your customer problem and needs."

He was equally direct about identity: "We are not a research lab converted to a consumer product now dabbling in enterprise. We are first and foremost an enterprise company that serves enterprise customers, and we evaluate our decisions within that lens. Which means, if we think building things from scratch is the right decision, that's what we will do. But if we think there are other alternatives out there in the market that serve our customers better, that's what we will do."

That pragmatism may be the release's most important signal. A well-capitalized American AI company with five years of model-building experience has concluded that the frontier of value no longer lies in pretraining, but in the last mile: post-training open weights, engineering the harness around them, and handing the CFO a dashboard. If Writer is right, the frontier labs' moat narrows to the workloads where quality genuinely justifies a sevenfold price premium — and for everything else, the winning model is the one somebody else paid to pretrain.

In an industry that has spent three years arguing about whose model is smartest, Writer is making a different wager: the enterprise AI race won't be won by the company with the best floating point numbers, but by the one that knows what to do with them.

Visa's president of technology, Rajat Taneja, walked the VB Transform 2026 audience through aiming Anthropic's Mythos at Visa's own payment network. The model stitched minor weaknesses into working exploit chains, and Visa open-sourced the harness that governed the hunt.

That's what it looks like when an enterprise has the engineering depth to act on what it finds. Most don't get there. Just over half, or 53%, of enterprises have already had an agentic security incident or near-miss. Sixty-five percent enforce agent permissions at runtime, yet only 18% isolate their highest-risk agents, and just 8% pair enforcement with isolation.

Leaning on provider-native controls to do the heavy lifting of agentic security just exacerbates that gap. The July wave of VentureBeat Pulse Research found that 92% of enterprises naming a primary security layer default to their hyperscalers and AI platform providers.

Six waves of research have been completed since January, surveying 440 qualified enterprise security respondents. The key takeaway: the containment gap between what enterprises need and what's getting done is growing wider, often unaddressed by enterprises whose agentic AI investments and futures are at risk.

The satisfaction data doesn't match the incident data

The research keeps showing enterprises rating the tools they know best at a higher score, even if those tools failed them or delivered mediocre results. Three findings from the raw data cut against that instinct, and each one says something about how young this market still is.

The enterprises that got hit rate their tools higher than the ones that didn't

Last month’s survey found that 46 enterprises reported a confirmed incident or near-miss, then went on to rate their satisfaction with their security tooling. Their average satisfaction was 4.39 out of 5. 30 of the 55 enterprises who experienced no incidents rated their security tooling at 4.13. Enterprises are rewarding any tool that saves them from a breach with a trust premium.

It’s a sure sign of a nascent market when brand positioning, marketing, or other means of persuading enterprises get easily superseded by saving a customer from a breach. Near-misses outnumber confirmed incidents 2-to-1 in both June and July, which means enterprises are catching problems at the edge. That edge catch is being interpreted as validation of both the security strategy and the tools acquired. Evident through seven months of data is how quick enterprise security leaders are to trust a new tool that identifies an intrusion or breach and defeats it before it gains access. VentureBeat believes the rescue itself is doing the marketing. The 4.13 average among never-hit enterprises shows the other side of the same effect. Tools that have never been seen working earn less trust, not more.

VentureBeat also found that of the 17 enterprises isolating their highest-risk agents, the 14 that rated their tooling average 4.00. Enterprises that do not isolate rate it 4.35. The enterprises closest to real security are the least satisfied with their tools — that dissatisfaction is what drives them toward the kind of engineering effort Visa put in.

Four of five enterprises that solved identity did not build isolation

49%, or 57 of the 116 enterprises surveyed in July, gave each agent its own scoped, managed identity. Just a month earlier, VentureBeat's June wave recorded 32% of enterprises having assigned per-agent identities. July’s 17-point jump in one month is the fastest single-month move this series has recorded. Despite these gains, 63% still report credential sharing somewhere in the fleet. Only 11 of those 57 also isolate.

That ratio explains why the containment gap keeps widening even as every headline control improves. Enterprises are treating identity and isolation as substitutes. They need to see the longer-term vision of each being integral to a platform-based, layered strategy. Two incidents VentureBeat has covered show why that distinction matters. A rogue AI agent at Meta passed every identity check before its March exposure was contained. And CrowdStrike CEO George Kurtz disclosed, at his RSAC 2026 keynote, a Fortune 50 agent that rewrote its own security policy using valid credentials. Giving an agent scoped credentials does not bound the blast radius when those credentials are misused. Sandboxing does.

The enforce-without-isolate population has a 58% incident rate

Fifty-three enterprises in July’s survey enforce scoped permissions at runtime but do not isolate. 31 of those 53 have already had an agent security incident or near-miss. That is 58%, five points above the 53% sample average. The enterprises living inside the containment gap are getting hit more often than the enterprises outside it.

Amy Chang, Cisco's head of AI threat intelligence and security research, presented findings on the Transform agentic security panel showing that when Cisco ran 6,986 multi-turn attacks against 15 flagship models, attackers who adapted across the conversation broke through up to 88.3% of the time. Single-turn red-teaming missed it. An adaptive attacker who defeats the guardrails lands inside whatever architecture sits behind them, and for 53 of the enterprises in this data, that architecture enforces but does not contain.

VentureBeat's Q1 Pulse Research tracked the same structural weakness earlier this year. Unauthorized tool or data access ranked as the most feared failure mode in every Q1 survey, growing from 42% in January to 50% in March. The April-May survey found only 4% of enterprises comfortable relying on model guardrails alone. Enterprises predicted they needed external controls, choosing to build enforcement over containment.

Enterprises built enforcement 35 points ahead of forecast. Isolation barely moved

The April-May survey asked 109 enterprises how they expected agent behavior to be controlled by the end of 2026, and 30% predicted runtime enforcement, 14% sandboxed execution, and 32% model-level guardrails. By July, 65% had built enforcement, more than double the prediction, while isolation reached 18%, roughly the rate they said it would. Enterprises built what was easy at twice the forecast and built what was hard at roughly the forecast. The April question asked for the primary control mechanism, single-select, while July's posture question allowed multiple selections, so the comparison is directional rather than exact.

Provider lock-in accelerated across all three quarters

Provider-native platforms already led usage in April-May, named by seven in ten enterprises describing their tooling. By June, 82% called one their primary agent security layer, and by July that share reached 92%, with OpenAI's guardrails leading at 44%, Microsoft Azure at 42%, Anthropic's managed-agent controls at 37%, and Google Cloud at 31%. Cloudflare at 11% and Cisco at 9% lead the dedicated specialists fighting over what remains. The identity tools most relevant to the credential-sharing gap are the smallest of all, with Microsoft Entra Agent ID at 7%, while Okta for AI Agents, non-human identity platforms, and runtime sandboxing tooling each sit at 3%. CrowdStrike CTO Elia Zaitsev told VentureBeat at RSAC 2026 that observing agent actions is a solvable problem but inferring intent is not. The provider bundle proves his point, solving observation while leaving containment unbuilt.

74% plan to replace tools they just rated a career-high satisfaction score

Satisfaction scores continue rising as enterprises gain more experience using tools and techniques to stop agentic AI-based attacks. Rising to 4.29 out of 5 in July from 4.2 in June, satisfaction is the highest reading in the series.

Despite the high satisfaction levels, 74% plan to replace their tools within 12 months, up from 59% in June. Only 26% intend not to change. VentureBeat believes early adopters are impatient to gain greater insights, and know what they don’t know about agentic security and resilience. Closing that knowledge gap is forcing churn into a market this young, and the raw answers resolve the paradox: 92% of enterprises naming a primary layer name a provider-native one. The 4.29 measures how easy it is to turn on a provider's guardrails. It does not measure how effective those guardrails are at preventing the incidents 53% of the same respondents already had.

The organizations closest to the threat are the least confident about it

In June, defenders led attackers 35% to 21%, but by July the split was 30-30, a dead heat. Among enterprises that have been hit, 39% now say attackers are ahead, against 20% of those that have not. Getting hit nearly doubles the pessimism but does not change the shopping. Just 10% of enterprises include any agent-identity product in their consideration set. Runtime sandboxing draws 6%, and those numbers hold regardless of incident history. VentureBeat covered the same blind spot in the June data. The label changed from agent security gap to containment gap, but the shopping did not.

Methodology

The posture question was answered by 93 of the 116 qualified July respondents, and the skippers are not hidden isolators. Twenty-three of the 25 who selected no posture option are organizations still evaluating agents, unsure of their status, or with no deployment plans, groups for which a security posture largely does not yet exist, so the 18% isolation figure reads on the enterprises actually running or piloting agents. April-May, June, and July are separate, independently fielded waves rather than a single tracked series, so month-over-month comparisons in this piece are directional rather than a measured trend. Base sizes for the cross-cuts differ by instrument. The identity question covers all 116 respondents, isolation covers the 93 who described a posture, and the satisfaction inversion of 4.39 versus 4.13 is computed on the 76 respondents who rated their tooling.

The bottom line

VentureBeat's cross-survey analysis of 573 enterprise respondents concluded in July that enterprises deployed AI agents ahead of the controls needed to manage them, and they did it knowingly. Three waves of security-specific data now show where the knowing stops.

Enterprises continue giving agents scoped identities and treating that as containment, but that assumption is false, and the incident data keeps proving it. In fact, 46 of 57 enterprises that solved identity did not build isolation. The enforce-without-isolate population's 58% incident rate is the clearest evidence that identity alone isn't enough. The containment gap will not close through satisfaction with what is easy. Whether enterprises build isolation and governed identity deliberately, or whether a confirmed incident that propagates does it for them, is the question the next wave will answer.

Elon Musk's company SpaceXAI, formerly known as xAI, has released Grok 4.6, its latest frontier AI model, with a focus on long-running agents, coding and knowledge work — and a pricing strategy designed to make those workloads cheaper to run.

The model scores 61 on the third-party Artificial Analysis Intelligence Index, surpassing the popular open weights Chinese model from Moonshot, Kimi K3, and tying rival OpenAI's GPT-5.6 Sol Max and improving five points over Grok 4.5 High. Anthropic's Claude Opus 5 and Fable 5 occupy the number one and two spots, respectively.

More consequential for enterprises evaluating AI agents, Grok 4.6 posts sizable gains over its predecessor across coding, terminal, knowledge-work and agent benchmarks while retaining an application programming interface (API) price starting at $2 per million input tokens and $6 per million output tokens, making it a mid-priced frontier model comparing leading options that are both proprietary and open source, globally, according to VentureBeat's analysis.

Still, that's less than half of what GPT-5.6 Sol costs over OpenAI's API in standard mode.

SpaceXAI says Grok 4.6 is available today in Grok Build, SpaceXAI's answer to Anthropic's Claude Code and OpenAI's Codex, which is available starting in the $30 per month SuperGrok plan.

It's also available in SpaceX's recent acquisition of the AI coding startup Cursor, and from partners including OpenRouter, Vercel and Cloudflare.

SpaceXAI is providing twice the included usage for Grok 4.6 in Cursor and Grok Build during the first week.

The release arrives only weeks after Grok 4.5, which SpaceXAI launched in July as a model targeting coding, agentic tasks and knowledge work, and one day after the launch of Grok Bot, a new system for assigning AI agents to complete designated tasks as virtual employees.

The bigger change is agent behavior, not just another benchmark point

SpaceXAI describes Grok 4.6 as being built specifically to stay on task across longer sequences of work, including researching unfamiliar topics, analyzing information, navigating codebases and converting product ideas into working applications.

The company says it subjected the model to a longer supplemental training run than Grok 4.5, using curated model-generated reasoning and technical data alongside engineering data and changes to its optimizer and training recipe. It then used Grok 4.5 to regenerate supervised fine-tuning trajectories across reasoning levels, agent harnesses, STEM, software engineering and knowledge work, filtering problematic trajectories with model-based checks.

Reinforcement learning also targeted agentic environments spanning general coding, knowledge work, kernel optimization, web development and computer-aided design.

That matters because enterprise AI deployments are increasingly moving beyond isolated prompt-and-response interactions toward agents expected to maintain state, operate tools, modify code and recover from problems across longer execution paths.

SpaceXAI says that during its testing, Grok 4.6 showed more self-testing and verification on longer trajectories, checking its own work before proceeding. It also reports stronger first attempts on interactive and visual projects than Grok 4.5. Those are company observations rather than independent guarantees of production behavior, but they indicate where SpaceXAI concentrated the model’s post-training work.

Grok 4.6 reaches the frontier, but does not sweep it

Grok 4.6's improvement over Grok 4.5 at this juncture of the AI model competition cannot be overstated.

According to Artificial Analysis, Grok 4.6 reaches an Elo score (human preference of head-to-head model outputs, adapted from chess) of 1,753 on GDPVal-AA v2, the benchmark measuring performance on real-world tasks like scheduling and diagramming, versus 1,526 for Grok 4.5, 1,728 for GPT-5.6 Sol Max and 1,741 for Fable 5 Max.

The coding results from SpaceXAI show a similar generational improvement but more competition at the frontier.

Grok 4.6 scores 69.9% on CursorBench v3.2, up from 66.7%, while Fable 5 Max reaches 70.5%. On DeepSWE v1.1, Grok rises sharply from 54% to 65.9%, but GPT-5.6 Sol Max leads at 73%. FrontierCode v1.1 Extended moves from 56.6% to 61.3%, compared with 60.6% for GPT-5.6 Sol Max and a leading 63.6% for Fable 5 Max.

Agent benchmarks tell much the same story. Grok 4.6 reaches 57.5% on APEX-Agents, a 10.4-point increase over Grok 4.5’s 47.1%, narrowly exceeding GPT-5.6 Sol Max’s 56.7% but trailing Fable 5 Max at 59.2%. On APEX-SWE, Grok 4.6 rises to 56.4% from 53.6%, while Fable 5 Max scores 58.8%.

Terminal-Bench v3.0 exposes a larger remaining gap. Grok 4.6 improves from 15.7% to 26%, but GPT-5.6 Sol Max and Fable 5 Max score 34.6% and 34.1%, respectively.

Two of Grok 4.6’s strongest results come from longer-horizon professional work. On AA-Briefcase it scores an Elo of 1,577, narrowly exceeding Fable 5 Max’s 1,574 and topping GPT-5.6 Sol Max’s 1,502. On Harvey LAB, Grok 4.6 reaches 15.8%, versus 12.9% for Grok 4.5, 11.3% for Fable 5 Max and 2.5% for GPT-5.6 Sol Max.

SpaceXAI notes an important methodological caveat: third-party scores in its table use the best self-reported or publicly available results. The comparison therefore should not be interpreted as a perfectly controlled four-model evaluation.

In other words, the evidence supports a substantial upgrade over Grok 4.5 more clearly than it supports across-the-board superiority over rival frontier models. Grok 4.6 wins several of the displayed evaluations while GPT-5.6 Sol Max and Fable 5 Max retain meaningful leads elsewhere.

Cost could be the more important enterprise benchmark

Artificial Analysis’ supplied evaluation adds another dimension: how much work the model performs for the money spent.

The testing places Grok 4.6 on its Intelligence-versus-Cost-per-Task Pareto frontier at a reported $0.84 per task — which actually makes it less of a bargain than its predecessor, Grok 4.5, and less economical than OpenAI's GPT-5.6 Luna, z.ai's GLM-5.2, and Meta's new Muse Spark 1.2, among other models.

Artificial Analysis also reports that Grok 4.6 completed its AA-Briefcase workloads in roughly 53 turns and about 0.5 billion input tokens on average, versus approximately 103 turns and 2 billion input tokens for Claude Opus 5 Max.

Those measurements do not prove that every production agent will use fewer tokens or finish twice as quickly. Agent costs depend heavily on harness design, prompts, tool calls, caching, retry behavior and the task itself. But they point toward an increasingly important enterprise metric: the cost of completing a workflow, rather than simply the cost of generating one million tokens. That distinction is central to SpaceXAI’s positioning.

The standard Grok 4.6 API starts at $2 per million input tokens and $6 per million output tokens, and SpaceXAI also offers a faster variant at twice the price.

The supplied API documentation adds an important caveat for long-context deployments. Grok 4.6 supports a 500,000-token context window, but prompts below 200,000 tokens are billed at $2 per million input tokens, $0.50 per million cached-input tokens and $6 per million output tokens. Once a prompt reaches 200,000 tokens, those rates rise to $4, $1 and $12 respectively, with the higher pricing applying to all tokens in that request.

That means enterprises should not extrapolate the $2/$6 headline pricing across the model’s entire context window when estimating total cost of ownership.

Artificial Analysis says the standard headline rates remain more than 60% below the competing frontier-model prices it cites for Claude Opus 5 and GPT-5.6 Sol. The practical savings will depend on how many tokens each model consumes to complete the same workload.

The Grok name carries considerable baggage and controversy, separate from the general AI skepticism

Performance and price may not be the only hurdles SpaceXAI faces in converting Grok 4.6's benchmark gains into enterprise adoption. The Grok brand arrives with an unusually visible history of safety and governance controversies — including extremist and antisemitic outputs, politically skewed responses, exaggerated praise of Elon Musk and, more recently, the use of Grok's image-generation capabilities to produce non-consensual sexualized imagery. For companies with strict compliance, brand-safety or responsible-AI requirements, that history could become a procurement consideration separate from the technical capabilities of Grok 4.6 itself.

The most notorious text-generation episode came in July 2025, when Grok produced antisemitic posts, praised Adolf Hitler and in some responses referred to itself as "MechaHitler." SpaceXAI's predecessor xAI subsequently said it was removing inappropriate posts and taking steps to prevent hate speech from being published by Grok.

Also in summer 2025, Grok began inserting references to an alleged "white genocide" in South Africa into answers to unrelated questions. xAI said an unauthorized modification to Grok's response software had directed the system to produce a particular response on a political topic while bypassing its normal review process. The company said the change violated its policies and subsequently pledged to publish Grok's system prompts and establish round-the-clock monitoring for problematic responses. The South African government has rejected claims that a genocide against white South Africans is taking place.

Grok's objectivity came under scrutiny again in November of the same year after the chatbot repeatedly produced implausibly flattering assessments of Musk. Among the examples reported at the time were claims placing Musk above elite athletes and historic intellectual figures. Musk said Grok had been manipulated through adversarial prompting into making "absurdly positive" statements about him. Whatever the underlying cause, the incident illustrated the reputational problem for an enterprise model whose outputs can become entangled with the public persona of the executive most closely associated with its developer.

The most serious controversy has involved image generation. In January 2026, U.K. regulator Ofcom opened a formal investigation into X after reports that the Grok account was being used to create and distribute undressed images of people and sexualized images of children. Ofcom said the material under examination could amount to non-consensual intimate-image abuse, pornography and child sexual abuse material. X subsequently said it had implemented measures intended to stop the Grok account from being used to create intimate images of people, but Ofcom said its investigation remained open.

The scrutiny extends beyond Ofcom. Britain's Information Commissioner's Office is investigating X and xAI over both the development and deployment of Grok, including whether personal data was handled lawfully and whether adequate safeguards existed to prevent harmful manipulated imagery. The European Commission, meanwhile, opened a separate formal investigation under the Digital Services Act examining X's management of systemic risks connected to Grok, including the dissemination of manipulated sexually explicit material.

Those investigations concern X and the earlier xAI organization rather than establishing a finding that the newly released Grok 4.6 API violates those laws. Nevertheless, they are unlikely to help SpaceXAI sell Grok to businesses.

SpaceX acquired xAI in February 2026, and the AI operation now markets itself as SpaceXAI, meaning Grok's newest models sit under a different corporate structure but retain the same consumer-facing brand.

There is no evidence in the material examined here that Grok 4.6 itself repeats the specific "MechaHitler," "white genocide," sexual-image or Musk-flattery incidents associated with earlier Grok deployments. But enterprise procurement teams rarely evaluate a model in isolation from its vendor and product history. For SpaceXAI, that means Grok 4.6 may have to demonstrate not only that it is cheaper or more capable than competing frontier models, but that the controls around it are sufficiently predictable for organizations that cannot afford their AI supplier to become a brand-safety event.

That continuity creates a potential adoption problem that benchmark tables cannot measure. Developers choosing a model for an internal coding agent may care primarily about price, latency and task completion. A bank, government agency, healthcare provider or consumer brand deploying the same model into customer-facing or regulated workflows may also have to consider vendor governance, content-safety controls, auditability and reputational exposure.

A model designed to be deployed, not just chatted with

Grok 4.6 supports text and image inputs with text output, function calling, structured outputs and reasoning, according to the supplied API specifications. Those specifications also list rate limits of 150 requests per second and 50 million tokens per minute, with API availability in us-east-1 and us-west-2.

Cursor’s launch announcement similarly characterizes Grok 4.6 as designed for long-running agents and ambitious interactive and visual work, giving developers immediate access to the model inside an established coding-agent environment rather than requiring them to build a new harness around the API first.

For enterprise buyers, that distribution may matter almost as much as another leaderboard result. Models increasingly compete not just on reasoning scores but on whether developers can place them inside existing coding, research and operational workflows without destabilizing those workflows or dramatically increasing inference costs, as well as incurring any blowback from associating with a controversial brand.

Grok 4.6 does not establish an uncontested performance lead. Its launch instead presents a different proposition: frontier-level intelligence, large improvements over the previous generation, stronger long-running agent behavior and relatively aggressive token economics.

The next test will be whether the efficiency Artificial Analysis observes on controlled agentic workloads carries into production. If Grok 4.6 can consistently complete long-running coding and knowledge-work tasks with fewer turns and fewer tokens, the model’s most important benchmark may ultimately be the enterprise inference bill rather than the leaderboard.

Skan AI, a startup that builds what it calls a "context graph of work" by observing how employees actually perform their jobs across enterprise software, has raised $63 million in Series C funding co-led by Cathay Innovation and Dell Technologies Capital, the company announced Wednesday.

Citi Ventures, Bloomberg Beta, State Farm Ventures, and Wipro Ventures also participated in the round, which brings the seven-year-old company's total funding to roughly $120 million. Alongside the raise, Skan is announcing the general availability of two new products — Skan AI Blueprint and Skan AI Agents — that, together with its existing Skan AI Intelligence offering, form a complete platform for discovering, modeling, and ultimately automating enterprise workflows.

The announcement lands at a moment of deep frustration in enterprise AI. Companies have poured billions into generative AI pilots, but the results have been dismal: Gartner research cited by the company finds that only 8% of enterprises have AI agents in production, and 95% of early implementations will require a complete redesign. Those figures echo an MIT report last year, covered by Fortune, which found that roughly 95% of enterprise generative AI pilots were failing to deliver measurable returns.

Avinash Misra, Skan's co-founder and CEO, believes the industry has misdiagnosed the problem. The models are fine, he argues. What they lack is an accurate picture of the businesses they are being dropped into.

"Everyone is obsessed with building a better driver," Misra told VentureBeat in an exclusive interview ahead of the announcement. "We think the bigger opportunity is building a better navigation system."

Why enterprise AI agents keep failing when they rely on official process documentation

The standard playbook for grounding AI agents — feeding them process documentation, standard operating procedures, and system logs — is built on a fiction, Misra argues. The way work is documented and the way work actually happens inside a large enterprise are two different things, and the gap between them is precisely where agents fail.

That gap is what sent Misra and co-founder Manish Garg down this path seven years ago, long before agents were a boardroom obsession. "Why is it so difficult for an organization, and a large enterprise especially, to understand how its own work actually gets done?" Misra said. "Why does it need to fly in McKinsey consultants for that?"

The question has only grown more consequential as enterprises race to operationalize AI. Frontier models arrive at the company door brilliant but blind, with no knowledge of the exceptions, decisions, handoffs, and institutional habits that define how a claims department or a compliance team actually operates. Every company now stuffing agents with documentation and logs, Skan contends, is discovering the same uncomfortable truth: the source data was never the whole story. And a source data problem cannot be fixed downstream.

Skan's answer is to go to the source itself. The company deploys observation technology on employee desktops that continuously watches how work moves across applications — the spreadsheet, the CRM, the email client, the 40-year-old mainframe — and abstracts those observations into a living model of the underlying business process.

"Think of it this way: if I were to share my screen here, and you were to observe my screen going from Excel sheet, CRM system, email client, in about two iterations you'd build a model of what I do," Misra said. "Except you couldn't do that at scale. You couldn't do it 24/7, and for 1,500 people like me. Now replace yourself with our technology."

How screen-level observation captures the work that never shows up in system logs

That framing also explains how Skan positions itself against process mining vendors like Celonis, which reconstruct workflows from the data trails left in backend systems. System logs, Misra argues, only capture completed transactions — not the messy human work that produced them.

"All backend data, by definition, is a committed state of work. Work is really what happens between those committed states," he said. "Eighty percent of what you're interested in, from an AI point of view, in execution of work, actually lies between those systems."

The screen, in Skan's view, is the one place where everything converges. "It brings together human agency, it brings together the entire application landscape, and it brings together the data that matters," Misra said. Two decades of user interface design have quietly buried enormous amounts of process knowledge in the space between a worker's eyes and their monitor; Skan's pitch is to bring that hidden layer back to the surface.

But watching, he insists, was never the hard part — a point aimed squarely at the incumbents who might be tempted to copy the approach. "The hard problem is not screen observation," Misra said. "The hard problem is abstraction of what you see on the screen — the intent extraction." A human watching a colleague's screen can instantly tell whether a jump back to step one means a new case or rework on an old one, because humans understand the signature of the work. Teaching a model to make that same judgment, statefully and at enterprise scale, is where Skan believes its seven-year head start lives.

The result is a context model that AI can reason over and act on — the raw material for the agents that now sit at the top of the company's product stack, and the foundation for everything else the platform does.

Walking the line between operational telemetry and workplace surveillance

An approach built on continuously watching employee screens invites an obvious objection, and it is not a hypothetical one. In June, Reuters reported that Meta scaled back an internal tool that tracked employee mouse clicks after workers raised concerns — a sign that even AI-forward companies are wary of the line between operational telemetry and surveillance.

Misra says he heard the objection before he wrote a line of code. When he first pitched the concept to Delphine Icart, then chief transformation officer at AXA Mexico, her reaction was blunt. "Delphine's first words to me were, 'This sounds like a great idea, but you are dead on arrival,'" Misra recalled. "'You are observing things that you shouldn't be observing — the privacy of my operators, and the sovereignty of my data on those screens.'"

That conversation, he says, shaped the architecture. Skan aggregates rather than individuates: the system surfaces statistical patterns across hundreds of workers performing the same process, not the behavior of any one of them. "We're not interested in what John is doing at 10 hours and 43 seconds," Misra said. "We are interested in what hundreds of Johns put together — what are the statistical and the semantic decisions that they are making in that business process?"

Organizations control what the technology can see through an opt-in scoping model — specific applications and URLs, nothing else — and the data Skan produces never leaves the enterprise firewall. A three-tier architecture sends only anonymized metadata to the cloud. Misra points to deployments approved by European works councils, among the most privacy-protective labor bodies in the world, as evidence the model holds up under scrutiny — and credits it for clearing security review at institutions where most AI tools cannot operate.

Whether aggregation fully defuses the concern is likely to remain contested. The same telemetry that reveals a broken process can, in principle, reveal an underperforming team, and Misra acknowledged that the technology has led some customers to reduce headcount in certain processes.

What $500 million in claimed customer value actually measures

Skan claims more than $500 million in cumulative customer value to date, a figure worth unpacking. Pressed on whether that represents realized savings or projections, Misra was direct that it is an envelope, not a bank balance.

"The number comes from the cumulative, across all our customers, of the quantified savings that we have brought to them — the savings that they have expected they would save," he said. "Now they are on the roadmap of recouping those savings through a variety of interventions," including process redesign, technology changes, and, increasingly, AI agents. In other words, $500 million is identified opportunity, some portion of which has been captured.

The more concrete evidence comes from individual deployments. At one top U.S. bank, according to the company, Skan observed 11.2 million context switches across 1,500 finance professionals and uncovered $37 million in operational friction. Turning those observations into agent-executable context cut cost per transaction by 32%, lifted throughput by 41%, and delivered $18 million in annualized savings.

Misra pointed to an anti-money-laundering operation at one bank where "60% of the cases are now being run by AI agents," adding that the results surprised even him: "The accuracy of those agents surpasses many times over the accuracy of humans. It's not just an argument of efficiency; it has also become an argument of quality." Among insurers, he said, Skan typically delivers roughly 25% productivity uplift in core claims processes; one customer doubled its case volume over the past year without adding a single claims specialist.

Skan's publicly referenceable customers include Unum, the $13.8 billion employee benefits provider, and Mitie, the U.K. facilities management company, whose chief technology and digital officer, Cijo Joseph, said Skan's technology "gives us unprecedented operational visibility that has dramatically accelerated our AI transformation." The company declined to share revenue but said it grew more than 300% year over year — for the second consecutive year — with net dollar retention around 150%, and now counts seven of the ten largest U.S. banks and a quarter of the Fortune 50 as customers.

Can AI models learn good work from imperfect employees?

Skan's thesis rests on observing how work actually gets done — which raises an uncomfortable question. Real employees make mistakes, take shortcuts, and entrench inefficiencies. What happens when the context graph faithfully encodes bad process?

Misra's answer reaches for the most famous precedent in modern AI. "Think for a moment what OpenAI did," he said. "OpenAI took the totality of the world's text and fed it into a transformer architecture, and semantic understanding emerged. OpenAI's model has seen bad language and has seen good language, and yet it is able to have semantic understanding."

Skan, he argues, does the analogous thing with work: treat business process execution as a language, where process steps, screen features, and handoffs stand in for words and sentences. Fed enough end-to-end executions, the model learns the full distribution of paths — efficient ones, slow ones, compliant ones — without assuming any single path is best. "The longest path may be the best path, because it is more compliant," Misra said. An organization then constrains the model along the axes it cares about, and the model returns the path that satisfies them.

"It is not record and play — and that's the fundamental difference between us and a lot of our competition, UiPath and so on," he said. "It is fundamentally creating an AI model that understands work, and then constraining that model."

He offered a concrete illustration of what that unlocks: at one large bank, Skan's telemetry continuously compares live case execution against a 600-page controls inventory, with agents that trigger alerts when cases miss required compliance steps — turning a document no human could hold in their head into a real-time enforcement layer. It is the kind of application that only becomes possible, Misra argues, once a model genuinely understands the work rather than merely replaying it.

The race to own the context layer of enterprise AI

Skan sits at the intersection of several crowded categories, and its answer to each competitor is a variation on the same theme: scope. Process mining vendors see only what the logs record. RPA incumbents replay tasks without understanding them. And the platform giants — ServiceNow, Salesforce, Microsoft — are shipping capable agents whose vision ends at their own walls.

"The context that these agents have access to is limited to ServiceNow, limited to Salesforce, whereas work spans processes across the board," Misra said. "Creating a customer entry is a task. To receive an email and decide whether a customer entry has to be created, or something else — that is the process, and that's what we are after."

The deeper strategic argument, and the one that seems to resonate with Skan's regulated customer base, is about differentiation in a world where every enterprise has access to the same frontier models. "If every insurance company, every bank had access to the same models, then the outcomes will asymptotically decay to the outcome of the model," Misra said. "Historically, you have competed and differentiated in the way you have organized work. That old word — process — now comes back as context for AI. But that context is protected by you. It's not part of the model."

That logic explains both the company's posture toward the model makers — "the more they are successful, the more power we have," Misra said, disclaiming any ambition to compete with them — and the Nvidia partnership featured prominently in the announcement. Skan runs on Nvidia AI Enterprise and NIM microservices, and Misra described growing demand for private appliances that can observe work, hold the context model, and execute agents entirely inside a customer's own infrastructure. It also fits the market's direction: venture investors surveyed by TechCrunch at the end of last year predicted enterprises would spend more on AI in 2026 but through fewer vendors — a consolidation that favors Skan's decision to ship discovery, intelligence, and agents as a single closed loop.

Misra argues that loop matters more, not less, as automation scales, because agents demand oversight in a way humans never did. "It is an irony of sorts," he said, "that you'll probably need much more observation and much more understanding of work in an automated way than you would with humans." The bet embedded in this round is that work context becomes foundational infrastructure for enterprise AI the way CRM became the system of record for customers — a comparison Cathay Innovation partner Simon Wu made explicitly, calling Skan "one of the defining platform companies of the next decade."

Misra put the stakes more simply. "You cannot retrieve context that you do not capture," he said. "The battleground is shifting from the smartest model to knowing how your company actually works — because everyone will have access to the smartest model."

The frontier labs, in other words, can keep their arms race for the better driver. Skan just raised $63 million on the conviction that the money is in the map.

Across 170 enterprises, AI infrastructure has moved decisively into production — two-thirds now run AI workloads live and three in 10 run them at scale — while the ability to account for what that infrastructure costs has not kept pace. Enterprises have quietly demoted cost in the buying decision: performance and GPU availability now outrank total cost of ownership, and reliability outranks price as the measure of success. That reordering is rational for teams under production pressure, but it lands on an uncomfortable fact — fewer than half can rigorously track what their AI compute costs, most GPUs still run at half capacity or less, and the next dollar is aimed at specialized clouds that fewer than one in twenty of them actually use.

This wave of VentureBeat Pulse Research examines enterprise AI infrastructure and compute: where organizations are in their deployment journey, what they run AI on today, how they buy and measure it, where the next investment is aimed, and — most revealingly — how well they can see the economics of the compute underneath it all.

This is an operational cohort. Two-thirds of enterprises (66%) have AI workloads running in production, and 29% describe AI in production at scale, with only 4% not yet running AI workloads at all. That maturity shows in the stack: the average enterprise runs three infrastructure platforms, with OpenAI (49%), Google Gemini (48%), Microsoft Azure (47%), and Google Cloud (42%) all present in roughly half of them. Asked to name one primary platform, Azure leads at 26%.

The most consequential shift is in how enterprises decide. Integration with the existing cloud and data stack remains the top selection factor at 40%, but performance — latency and throughput — has climbed to second at 35%, and access to GPU availability to third at 24%, both ahead of total cost of ownership at 22%. The same ordering governs measurement: uptime and reliability is the primary success metric for 51% of enterprises and developer productivity for 39%, ahead of cost per million tokens at 31%. Enterprises under production pressure are buying and measuring for speed and availability, and have moved cost down the list.

That would be unremarkable if the economics were under control, but they're not. Among the 155 enterprises that operate their own GPUs, 69% report utilization of 50% or less and only 23% clear the halfway mark; 12% do not measure utilization at all. Fewer than half (47%) rigorously track what their AI compute costs and returns, and even among enterprises running AI in production at scale that figure only reaches 56%. Value for money is the weakest of three satisfaction scores at 3.87, against 4.14 for overall satisfaction — the softness landing precisely on the dimension hardest to judge without measurement.

The next round of spending points away from the current stack. AI-specialized clouds are the top planned evaluation area at 44% and carry the strongest net momentum of any infrastructure approach (+36), yet CoreWeave and Lambda each registers at 3.5% of current usage and the rest of the neocloud field sits below 3%. Non-Nvidia accelerators draw 39%. And 62% of enterprises intend to switch or add a provider within 12 months — though the consideration set is dominated by the same incumbents they already run.

Methodology

VentureBeat fielded this survey as part of its ongoing Pulse Research series, this one focused on enterprise AI infrastructure, compute, and inference economics. Responses are filtered to organizations with more than 100 employees (n=170; the survey’s smallest size band, 1–100 employees, is excluded), drawn from a single July 2026 wave. Because this is one wave rather than a pooled multi-month sample, the report reads cross-sectionally and does not infer month-over-month trends; all figures are drawn from the July fielding only. Several questions were multiple-select, so those shares can sum to more than 100%.

By organization size this wave reaches further up-market than the mid-market skew this series usually carries: 251–1,000 employees (28%) and 1,001–5,000 (25%) lead, with 10,001+ (19%), 101–250 (15%), and 5,001–10,000 (12%) filling out the rest — meaning 57% of respondents sit above 1,000 employees. By role it spans managers (48%), individual contributors (27%), the C-suite (12%), and VPs and directors (9%); on purchasing authority it is buyer-credible, with 39% final decision-makers and another 43% recommenders or influencers for AI solutions. Technology/Software is the largest industry at 35%, followed by Manufacturing (14%), Financial Services (12%), and Healthcare/Life Sciences (9%).

At 170 respondents the sample is large enough to read directionally with reasonable confidence, but it should still be treated as a directional signal rather than a precise measurement; it is self-selected and is not a probability sample. It is best read as the view from organizations actively building and operating AI infrastructure rather than from the largest hyperscale operators.

Finding 1: Two-thirds are past the pilot

Three in 10 now run AI in production at scale

We asked where organizations sit in their AI deployment journey. This cohort has largely moved beyond experimentation.

Two-thirds of enterprises (66%) have AI workloads running in production, and 29% describe AI in production at scale. Only 30% remain in proofs of concept and just 4% have not started. This is a materially more operational sample than this series has typically drawn, consistent with its up-market composition — 57% of respondents sit above 1,000 employees.

That maturity is the frame for everything that follows. The infrastructure decisions in this report are being made largely by organizations with production workloads and real bills, not by teams still sizing a pilot. It explains the reordering of buying criteria in Finding 5, where performance and availability displace cost — the priorities of teams running live systems. It also raises the stakes on Findings 6 and 7: an enterprise that cannot measure utilization or cost during experimentation has a planning problem, while one that cannot measure them in production at scale has an operating one.

Finding 2: The stack is hyperscaler-and-API, three platforms deep

The specialized GPU clouds still barely register

We asked which providers and platforms enterprises currently use to run their AI, and which one they treat as primary. The answer remains the incumbents — several of them at once.

The current stack is hyperscaler-and-API, and it is plural: enterprises name three platforms on average. The general-purpose clouds and the major model APIs account for essentially all current deployment, with four platforms — OpenAI, Gemini, Azure, and Google Cloud — each presents in more than four of every 10 enterprises. Asked to pick one primary platform, Microsoft Azure leads at 26%, with Google Cloud second at 19%; the model providers together take 35% of primary status when OpenAI (14%), Gemini (14%), and Anthropic (8%) are combined.

The specialized “neocloud” GPU providers that dominate AI-infrastructure headlines remain marginal in practice. CoreWeave and Lambda each appear in 3.5% of stacks, Baseten in 3%, and Crusoe, Nebius, Fireworks, Together, and Anyscale each at or below 2%. Combined, they are named as the primary platform by 1% of enterprises. Meanwhile 13% run a custom open-source self-managed stack and 9% operate their own GPU clusters — both larger footprints than the entire specialized-cloud category. That contrast is what makes the evaluation intentions in Finding 3 worth reading closely.

A note on reading these shares: As described in the methodology section, this sample is self-selected and this question counted every provider a respondent uses — an average of 3.0 selections each — so the figures measure presence in the stack rather than spending or primary status. The separate primary-platform question is the better guide to where the center of gravity sits. A sample built this way will show a different provider mix than a spend-weighted census of the broader market; read these shares as a portrait of what this AI-active cohort runs today, and treat gaps against industry-wide market share estimates as a property of the sample rather than a contradiction of either.

Finding 3: The next dollar goes to infrastructure they don't yet run

AI-specialized clouds top the evaluations list and carry the strongest momentum

We asked where enterprises plan to evaluate AI infrastructure over the next 12 months, and whether they expect to do more or less with each category of infrastructure. Both answers point away from the stack they run today.

Here is the report’s sharpest tension, and it is the same one this series has now recorded across successive waves. The single most-cited planned evaluation area — AI-specialized clouds, at 44% — is the category that 3.5% of these enterprises actually use (Finding 2). Nearly four in 10 (39%) intend to evaluate non-Nvidia accelerators, a quarter next-generation Nvidia silicon, and even decentralized compute networks draw 18%.

The direction-of-travel question corroborates it rather than merely repeating it. Asked whether they expect to do more, less, or about the same with each approach, enterprises put specialized AI clouds at the highest net momentum (+36, with 42% doing more against 6% doing less), ahead of inference APIs (+34) and hyperscalers (+30). On-prem and co-located infrastructure is the laggard at +5, the only category where a substantial share — 22% — report pulling back. Every off-premises approach is net-expanding; the specialized clouds are expanding fastest from the smallest base.

Read against current usage, this is not incremental adjustment. It is the leading edge of a re-platforming that enterprises have been signaling for several waves and have not yet executed. The gap between a 44% evaluation rate and a 3.5% usage rate is the single widest intent-to-action spread in this dataset, and how it resolves — whether the neoclouds convert evaluation into deployment, or whether the hyperscalers absorb the demand with their own AI infrastructure — is the open question of the category.

Finding 4: Six in 10 plan to move, mostly among the incumbents

High churn intent, but the consideration set is the stack they already run

We asked whether and when enterprises plan to switch or add an infrastructure provider, and which providers they are considering.

For a category as foundational as compute, this is a substantial amount of intended movement: 62% of enterprises intend to switch or add a provider within 12 months, and 29% within the next quarter alone. Only 39% plan to stand still.

Where that interest points is the more useful signal. The providers drawing the most switching consideration are the ones enterprises already run — OpenAI and Google Cloud (29% each), Microsoft Azure (28%), Gemini (25%), Anthropic (16%), Oracle Cloud (14%), and AWS (13%). The specialized clouds that top the evaluation list in Finding 3 draw far less concrete switching consideration: CoreWeave 4%, Lambda 3.5%, and the remainder at or below 2%. A further 8% are evaluating with no shortlist yet.

The two findings are not in conflict; they operate on different clocks. The neocloud interest in Finding 3 is a 12-month evaluation thesis about where AI compute should eventually run. The switching in the next quarter is mostly incumbents trading share and enterprises consolidating spend among providers they already hold contracts with. Vendors reading the 44% evaluation figure as near-term pipeline should weigh it against a 4% consideration rate.

Finding 5: Performance overtakes cost, in buying and in measurement

Total cost of ownership falls below latency and GPU availability

We asked what matters most when enterprises select an AI infrastructure provider, and what they treat as the primary measure of success once it is running. Both answers have moved away from price.

Integration with the existing stack remains the top selection factor at 40%, which is consistent with a cohort running three platforms and unwilling to add a fourth that does not fit. What has changed is everything below it. Performance sits second at 35% and GPU access and availability third at 24%, both ahead of total cost of ownership at 22%. Fine-grained autoscaling draws 18% and cost per million tokens 16% — no longer the outlier it once was in this series, but still last.

Measurement follows the same logic. Uptime and reliability is the primary success metric for 51% of enterprises, well ahead of developer productivity and deployment speed (39%), cost per million tokens (31%), latency (27%), and throughput (25%). Taken together, the operational metrics dominate the economic one by a wide margin.

This is a coherent posture for the production cohort in Finding 1 — teams running live workloads care first about whether the system stays up and how fast they can ship on it. But it sits uneasily beside Finding 7. Total cost of ownership has been demoted to fourth as a buying criterion at exactly the moment when 53% of enterprises still cannot rigorously track what their compute costs. The uncomfortable reading is that cost has fallen down the list partly because it remains the hardest thing in the stack to see, and criteria that cannot be measured tend to lose to criteria that can.

Finding 6: The GPUs run warmer, but most still run cold

Roughly seven in 10 GPU operators report 50% utilization or less

We asked what share of their GPU capacity enterprises actually utilize. Figures here are reported on the 155 enterprises that operate their own GPUs; 15 consume exclusively via API and run none.

The compute already in place runs cold, though less so than this series has recorded before. Roughly seven in ten GPU-operating enterprises (69%) report utilization at or below half capacity, with the 26–50% band alone accounting for 46%. About a quarter (26%) run at 25% or below. Against that, 23% now clear the 50% mark — a meaningful efficient minority rather than a rounding error.

The remaining 12% who do not measure utilization at all are the more troubling number, because they are invisible in both directions: they cannot claim efficiency and cannot detect waste. And utilization does not improve with maturity in the way one might expect — among enterprises running AI in production at scale, 22% clear the 50% mark, statistically indistinguishable from the 24% among everyone else. Scale is not, by itself, producing better-utilized fleets.

Idle accelerators are expensive accelerators, and this remains the clearest single measure of the gap in this report: enterprises are planning to evaluate specialized clouds and next-generation silicon (Finding 3) while the capacity they already own sits substantially unused. The efficiency headroom in the current fleet is large, and for one in eight enterprises, entirely unmeasured.

Finding 7: Fewer than half can account for what they spend

Rigorous cost tracking reaches only 56%, even among at-scale operators

We asked whether enterprises can quantify the cost and return of their AI infrastructure spend, and how satisfied they are with what they run. Confidence in the ledger still lags the spending.

Measurement trails money. Fewer than half of enterprises (47%) rigorously track the cost and return of their AI compute; the majority track only partially (39%), cannot quantify it yet (15%), or have not prioritized it (6%). Maturity helps but does not solve it: among enterprises running AI in production at scale, rigorous tracking reaches 56%, against 43% for everyone else. Even in the most operationally advanced segment of this sample, more than four in ten cannot account precisely for what their AI compute costs or returns.

Satisfaction with current infrastructure is moderately positive and tellingly uneven. On a five-point scale, overall satisfaction averages 4.14 and ease of implementation 4.04, while value for money trails at 3.87 — the softness landing on the one dimension that requires measurement to assess. Enterprises are, in effect, expressing dissatisfaction with an economic relationship most of them cannot yet quantify.

Read with Finding 5, the picture is self-reinforcing rather than merely inconsistent. Cost has slipped to fourth among buying criteria while remaining the least visible property of the stack, and the least visible property is the one enterprises rate lowest. Better instrumentation would not necessarily change what enterprises buy — but it would let them know whether the trade they are making for performance and availability is a good one.

Finding 8: The memory frontier is still unclaimed

Dell and Nvidia lead a scattered field, and one in five has no view

We asked how enterprises would address the emerging constraint in large-scale inference — the shift from GPU compute to memory, specifically KV-cache capacity. The field remains early and fragmented.

The memory frontier is real but barely governed. Dell leads at 24% and Nvidia follows at 21%, with the remainder scattering across open-source tooling (12%), model-level efficiency techniques such as MLA and quantization (11%), and a long tail of storage vendors each in low single digits. No approach commands anything close to a majority, and the two leaders together account for less than half the field.

Most telling is that roughly one in five enterprises (19%) either do not recognize the constraint (7%) or have not begun to address it (12%). For a shift that will reshape inference cost and architecture, this is an early and unsettled market. It is also consistent with the measurement gap in Finding 7 — enterprises that cannot yet quantify what their current compute costs are in a poor position to anticipate which constraint will drive that cost next. The memory bottleneck is arriving while most of this cohort is still working to see the one in front of it.

The bottom line: Buying for speed, blind on cost

Organizations with more than 100 employees have moved AI infrastructure into production — two-thirds run live workloads, three in ten at scale — and their buying behavior has matured accordingly. They run three platforms on average, select on integration and performance, and measure success on uptime and developer velocity. For teams operating live systems, that is the right set of priorities.

What has not matured is the accounting. Total cost of ownership has fallen to fourth among selection criteria and cost per million tokens sits last, at the same moment that 53% of enterprises cannot rigorously track what their compute costs, 69% of GPU operators run at half capacity or less, and 12% do not measure utilization at all. Value for money is the lowest-rated attribute of the infrastructure they run — a judgment most of them are making without the instrumentation to support it. Cost has not become unimportant; it has become invisible, and the buying criteria have quietly reorganized around what can actually be seen.

Meanwhile the next round of spending points past the current stack. Specialized AI clouds are the top evaluation target at 44% and carry the strongest net momentum of any approach, against a 3.5% usage rate and a 4% near-term switching consideration — the widest intent-to-action spread in the data. Non-Nvidia accelerators draw 39%. And the constraint after this one, the shift from compute to memory in large-scale inference, is unrecognized or unaddressed by one enterprise in five.

At 170 respondents in a single July wave, reaching further up-market than this series typically does, this is a directional read — but the direction is consistent. Enterprises have become good operators of AI infrastructure and have not yet become good accountants of it. The open question for later waves is whether the instrumentation catches up before the re-platforming arrives, or whether enterprises buy the next layer of compute as blind to its economics as the last.

Based on survey responses from 170 qualified enterprise respondents (100+ employees), drawn from a single July 2026 wave. This sample is self-selected and directional rather than a precise measurement, and reads cross-sectionally with no month-over-month trend claims. Respondents include managers, individual contributors, C-suite, and VPs/directors, with purchasing authority weighted toward decision-makers and recommenders, across technology, manufacturing, financial services, healthcare, and other industries. Note: Figures for the switching-timeline, GPU-utilization, and cost-tracking questions are reported as a percentage of unique respondents rather than selections; individual categories for these three questions may sum to more than the reported total.

Across 116 enterprises, agents are in production and so are the incidents: A majority have already had a confirmed agent security event or a near-miss. Two-thirds of enterprises enforce scoped permissions at runtime. Barely one in five isolates its highest-risk agents, making containment the weakest layer in the stack precisely as autonomy scales. Credential sharing persists across nearly two-thirds of agent fleets, and 53% have already had a confirmed agent security event or near-miss, contributing to a growing lack of confidence in agentic security.  Security stacks remain overwhelmingly borrowed from model providers and hyperscalers, and confidence has slipped. Today, as many enterprises now believe AI-armed attackers are ahead of their defenses as believe the reverse.

This wave of VentureBeat Pulse Research examines how enterprises secure their AI agents: what tooling they run, how they manage agent identity and isolation, what has already gone wrong, how much they spend, and whether they believe their defenses are keeping pace with AI-enabled attackers.

Only 18% of enterprises isolate their highest-risk AI agents, even as 65% of enterprises enforce scoped permissions at runtime and 56% monitor and log agent activity. The gap between what enterprises watch and what they contain is the central finding of this wave of VentureBeat Pulse Research.     

More than half of enterprises (53%) have agentic AI systems in production today, and another 27% are piloting or running a limited rollout. The agentic security incidents are arriving with them: 53% of organizations have already had an agent security event, with 19% confirming an incident and 38% having identified a near-miss that was caught before it caused harm.

The central finding is a containment gap. Enterprises have built the controls that watch and permission agents but not the one that bounds the damage when those fail. Among enterprises describing their security posture, 65% enforce scoped identities and permissions at runtime and 56% observe and log agent activity, yet only 18% isolate high-risk agents in sandboxes. Even among enterprises running agents in production, isolation is enforced just 21% of the time, and just 8% pair enforcement with isolation. That ordering is backward from a defense-in-depth standpoint. From SOC teams to CISOs, security teams know that observation tells you what happened and enforcement tries to prevent it, but isolation is what limits the blast radius when prevention fails.

Identity has improved without being solved. 49% of enterprises say each of their agents has its own scoped, managed identity, but 63% report credential sharing somewhere in the agent fleet, and only 29% describe a fleet with scoped identities and no sharing anywhere. The security stack doing this work remains overwhelmingly hyperscaler or model provider-native: OpenAI’s guardrails (44%), Microsoft Azure (42%), Anthropic’s managed-agent controls (37%), and Google Cloud (31%) lead, and 92% of enterprises naming a primary security layer name a hyperscaler/model provider-native one.

Two things have shifted against the comfortable picture. Confidence has slipped, with 30% now saying AI-armed attackers are ahead of their defenses, exactly as many as say their defenses are ahead. And churn intent is the highest this series has recorded, with 74% planning to adopt, add, or replace agent security tooling within twelve months, despite satisfaction scores at a series high of 4.29 out of 5. Enterprises are more satisfied than ever with a stack they are more determined than ever to replace.

Methodology

VentureBeat fielded this survey as part of its ongoing Pulse Research series, this instrument focused on enterprise agent security — the tooling, identity, isolation, and enforcement controls organizations use to secure autonomous AI agents. Responses are filtered to organizations with more than 100 employees (n=116; the survey’s smallest size band, 1–100 employees, is excluded), drawn from a single July 2026 wave. Because this is one wave rather than a pooled multi-month sample, the report reads cross-sectionally and does not infer month-over-month trends; all figures are drawn from the July fielding only. Several questions were multiple-select, so those shares can sum to more than 100%.

By role the sample is senior and buyer-credible: 44% are final decision-makers for AI purchases and another 38% recommenders or influencers. Managers (36%), individual contributors (27%), VPs and directors (18%), and the C-suite (16%) make up the seniority mix. By organization size the sample is mid-market-weighted with a meaningful enterprise tail: 101–250 (34%) and 251–1,000 (23%) employees lead, with 1,001–5,000 (18%), 10,001+ (17%), and 5,001–10,000 (7%) above them. Technology/Software is the largest industry at 38%, followed by Healthcare/Life Sciences (11%) and Financial Services (10%).

Three questions require a base note. Two questions were asked only of enterprises with agents live or piloting. Posture figures (observe / enforce / isolate) are reported on those 93 respondents, and primary-security-layer figures on the 92 of them who named a layer. The 23 respondents outside this base are those still evaluating, without plans, or unsure — organizations for which an agent security posture would not yet apply. And several multiple-select questions permitted overlapping answers where one was intended — identity (33 respondents selected more than one pattern), arms-race assessment (23), budget share (10), and incidents (9) — so those are computed at the respondent level and the overlap is described where it matters. Satisfaction ratings are computed on the respondents who answered each rating question; the overall satisfaction score reflects 76 of the 116 qualified respondents.

At 116 respondents, the sample supports directional reads but not precise measurement; it is self-selected and is not a probability sample. It is best read as the view from organizations actively standing up agent security rather than from the largest operators.

Finding 1: Agents are in production, and so are the incidents

A majority have already had an agent security event

We asked whether organizations run agentic AI in production, and whether they had experienced an agent security incident — a confirmed breach, or a near-miss caught before harm.

Agents have moved into production for this cohort. More than half of enterprises (53%) run agentic AI systems live today, another 27% are piloting or running a limited rollout, and only 3% have no plans in the next twelve months. The security exposure has scaled with the deployment: 53% of organizations have already had an agent security event, 19% a confirmed incident and 38% a near-miss caught before it caused harm.

That the near-misses outnumber confirmed incidents two to one is worth reading carefully. It means enterprises are catching problems, but catching them close to the edge — and a near-miss is a control that worked once, not a control that will work every time. The controls examined in the rest of this report, particularly the identity and isolation gaps in Findings 2 and 3, are what determine whether the next near-miss stays a near-miss.

One pattern from earlier waves does not replicate here. Organization size makes no reliable difference to exposure: enterprises above 1,000 employees report an incident or near-miss at 47%, against 57% among those between 101 and 1,000 — a difference well inside sample noise, and pointing the opposite direction from the size gradient this series has previously recorded. In this wave, what separates the hit from the not hit is not headcount.

Finding 2: Identity is improving — and still shared

Half give agents scoped identities; two-thirds still share credentials somewhere

We asked how enterprises manage the identity of their AI agents — whether each agent has its own credentials, or agents share them. Respondents could describe more than one pattern across the fleet.

Per-agent identity is now the most-cited pattern: 49% of enterprises say each agent carries its own scoped, managed identity, the precondition for least-privilege access and clean attribution. That is real progress on the control this series has repeatedly identified as the structural weakness beneath agent incidents.

But the answers overlap, and the overlap is the finding. Thirty-three respondents described more than one identity pattern across their fleet, and rolled together at the respondent level, 63% of enterprises report credential sharing somewhere — either agents mostly running on shared API keys and borrowed human or service-account credentials (37%), or a mixed fleet where some agents are scoped and many are not (34%). Only 29% describe a fleet with scoped identities and no sharing anywhere at all. Among enterprises with agents in production, 60% report per-agent identity, so the improvement is concentrated where the agents actually are — but so is the residual sharing.

The consequence is unchanged by the improvement. Where credentials are shared, an over-permissioned or compromised agent acts with far more reach than intended, and post-incident forensics cannot cleanly establish which agent did what. Half a fleet with scoped identities still has the blast radius of the half without. Non-human identity remains the largest unfinished piece of enterprise agent security, and as Finding 8 shows, it is still almost entirely absent from what enterprises are shopping for.

Finding 3: Isolation is the control nobody builds

Two-thirds enforce at runtime; fewer than one in five sandbox

We asked what an organization’s agent security posture looks like in practice — whether they observe, enforce, isolate, or some combination. The control that bounds damage is by far the least common. Figures are reported on the 93 respondents who described a posture.

This is the containment gap, and it is the widest structural gap in the report. Enforcement and observation are now common — 65% enforce scoped permissions at runtime and 56% monitor and log agent activity — while isolation sits at 18%. Only 8% of enterprises run both enforcement and isolation together, the posture that both prevents and contains.

Deployment maturity is a better predictor than the aggregate figures suggest. Isolation reaches 21% among enterprises with agents fully in production, compared with 13% among those still piloting — a meaningful gap that tracks maturity rather than exposure. Among enterprises that report credential sharing in the fleet, the group with the widest potential blast radius per Finding 2, isolation reaches 15%. The organizations with the most exposure are not meaningfully more likely to have built the control that bounds it.

The ordering is backwards from a defense-in-depth standpoint. Observation tells you what happened after the fact. Enforcement tries to stop it. Isolation is what limits the damage when enforcement fails — and enforcement will sometimes fail, which is the entire premise of the near-misses in Finding 1. An agent fleet that is watched and permissioned but not boxed in is precisely the configuration in which a single control failure propagates across systems. Enterprises have built the first two layers of the model and largely skipped the third.

Finding 4: Security still runs on borrowed, provider-native controls

Nine in 10 name a model provider or hyperscaler as their primary layer

We asked which agent security tooling enterprises use, and which is their primary layer. The answer continues to favor the model providers and hyperscalers over the dedicated security vendors.

Enterprises secure agents with tools that came bundled with their models and clouds. OpenAI’s guardrails lead at 44%, followed closely by Microsoft Azure (42%), Anthropic’s managed-agent controls (37%), and Google Cloud (31%). Asked to name a single primary security layer, 92% of those who answered named one of these provider-native offerings, with Azure (27% of answerers) and Anthropic (26%) leading.

The purpose-built agent-security category is no longer at zero, but it remains marginal. Cloudflare (11%) and Cisco (9%) lead the specialists, with CrowdStrike, Palo Alto, Zenity, Check Point’s Lakera, HiddenLayer, F5, and SentinelOne each between 1% and 7%. The identity specialists most directly relevant to Finding 2 are the smallest of all: Microsoft Entra Agent ID at 7%, Okta for AI Agents at 3%, and non-human identity platforms at 3%. Dedicated runtime sandboxing tooling — the control missing in Finding 3 — is in place at 3%.

A note on reading these shares: As described in the methodology section, the respondent sample is self-selected, and the usage question counted every vendor or approach a respondent has in place — so the figures measure presence in the security stack rather than spending or exclusivity. Individual vendor percentages therefore carry all the usual sample caveats. The structural pattern is the durable part: provider-native and hyperscaler controls lead by a wide margin, and dedicated agent-security specialists remain in single digits. Read the individual shares loosely and the pattern with confidence.

Finding 5: Satisfaction is at a series high — and so is churn intent

Enterprises rate their tooling 4.29 of 5 and three-quarters plan to replace it

We asked how satisfied enterprises are with their current agent security tooling, and whether they plan to adopt a new, additional, or replacement solution within twelve months. The two answers do not sit comfortably together.

Satisfaction with agent security tooling is the highest this series has recorded — 4.29 out of 5 for both overall satisfaction and ease of implementation, with value for money close behind at 4.11. That is a striking set of scores for a stack that is mostly borrowed provider guardrails, given that a majority of the same enterprises have already had an incident or near-miss and fewer than one in five isolates high-risk agents.

The purchase intentions tell the other half of the story. Three-quarters (74%) plan to adopt, add, or replace agent security tooling within 12 months, and 30% within the next quarter alone — higher churn intent than this series has previously seen in this category. Only 26% intend to stand pat. Enterprises are simultaneously more satisfied with their tooling and more determined to change it than at any prior reading, which suggests the satisfaction rests on the convenience and low friction of provider-native controls rather than on demonstrated containment. It is comfort with what is easy, not confidence in what is sufficient.

Finding 6: Budgets are finally moving

A third now spend more than a tenth of the security budget on agents

We asked what share of the security budget enterprises allocate to securing AI agents. The allocation has grown, though it remains a modest slice.

Agent security spending is still a slice rather than a pillar, but it is a growing one. The most common allocation remains 6–10% of the security budget (44%), and roughly a third of enterprises (35%) now devote more than a tenth — a meaningful funded minority. Just over a quarter (28%) spend 5% or less.

Read against Findings 1 through 3, the budget looks like a lagging but responsive indicator. A majority of enterprises have had an incident or near-miss, credential sharing persists across two-thirds of fleets, and fewer than one in five isolates high-risk agents — gaps that a 6–10% allocation is unlikely to close quickly. The enterprises spending above a tenth are the ones with the resources to build scoped identity and isolation controls rather than adopt whatever their model provider ships, and whether that minority grows is a reasonable leading indicator for whether the containment gap narrows.

Finding 7: The arms race has tilted

As many say attackers are ahead as say their defenses are

We asked how enterprises assess the balance between their AI-enabled defenses and AI-enabled attackers. Confidence has slipped into an even split.

Enterprises are no longer net-optimistic about the contest. Exactly as many say AI-armed attackers are ahead of their defenses (30%) as say their defenses are ahead (30%), with another 33% calling it roughly even and 24% saying it is too early to tell. Taken together, 63% rate the balance as even or worse.

Experience is what drives the pessimism, and the relationship is statistically clear. Among enterprises that have had a confirmed incident or near-miss, 39% say attackers are ahead; among those that have not, 20% do — a gap large enough to be unlikely to arise by chance in a sample this size. Getting hit does not just change what enterprises buy; it changes how they read the contest. The organizations closest to the actual threat are the least confident about it.

That assessment sits uneasily beside the series-high satisfaction of Finding 5. Enterprises rate their tooling 4.29 out of 5 while a clear majority believe it is, at best, holding even against an adversary that is also compounding with AI. An even race is not a comfortable place to be, and the group that has actually been tested rates it worse than even.

Finding 8: A reshuffle is coming — but identity still isn’t on the list

Incidents drive urgency; the control they implicate draws 10% interest

We asked which agent security solutions enterprises are considering. The consideration set has broadened, but not in the direction the incident data points.

Incidents start the buying cycle. Among organizations that have had a confirmed incident or near-miss, 38% plan to adopt, add, or replace agent security tooling within the next ninety days, against 22% of organizations with no incident; after a confirmed incident specifically the figure reaches 41%. Experience remains the strongest predictor of urgency in this data, as it is of pessimism in Finding 7.

The consideration set still leans provider-native — OpenAI (38%), Microsoft Azure (37%), Anthropic (35%), and Google Cloud (28%) lead — though the dedicated security vendors now draw meaningful early interest: Cisco (10%), Cloudflare (9%), Zenity and CrowdStrike (8% each), and Palo Alto, Check Point’s Lakera, and open-source guardrails (6% each). For most of the specialists that is more forward interest than current footprint.

What the shopping still does not include is the identity layer. Just 10% of enterprises include an agent-identity product — Okta for AI Agents, Microsoft Entra Agent ID, or a non-human identity platform — anywhere in their consideration set. Among the enterprises that both share credentials and have already been hit, the group with the most direct evidence that the control matters, identity consideration is no higher: roughly one in ten. Runtime sandboxing tooling draws 6%. The two controls most directly implicated by the incident data, identity and isolation, are the two least present in the purchase plans — the same blind spot this series recorded in the prior wave, unchanged despite a year of incidents.

The bottom line: A security gap that prevention alone won’t close

Organizations with more than 100 employees have put agents into production — 53% run them live today — and the incidents have arrived alongside them, with a majority already reporting a confirmed event or near-miss. On the controls, the picture is genuinely mixed rather than uniformly poor: nearly half now give each agent its own scoped identity, two-thirds enforce permissions at runtime, and a third devote more than a tenth of the security budget to agents. Enterprises are building agent security in earnest.

What they are not building is containment. Fewer than one in five isolates high-risk agents, only 8% pair enforcement with isolation, and among enterprises running agents in production isolation reaches just 21%. Credential sharing persists across 63% of fleets, so the blast radius that isolation would bound remains wide. The stack doing this work is 92% provider-native by primary layer, and the specialists built for exactly these gaps sit in single digits. The result is an architecture optimized to prevent and observe, with almost nothing in place for the case where prevention fails — which is the case the near-misses in Finding 1 describe.

The uncomfortable pairing is confidence with exposure, and it has sharpened. Satisfaction is at a series high of 4.29 out of 5, yet 63% rate the contest against AI-armed attackers as even or worse, 30% say attackers are ahead outright, and 74% plan to replace tooling they just rated highly. Enterprises that have actually been hit are markedly more pessimistic and markedly more urgent — and still not shopping for identity or isolation, the two controls their incidents most directly implicate.

At 116 respondents in a single July wave this is a directional read, weighted toward the mid-market — but the direction is clear: agent deployment is running ahead of agent containment, and the gap is not in what enterprises watch or permission but in what happens when those controls fail. The containment gap will not be closed by a better provider guardrail. The open question for later waves is whether enterprises build isolation and governed identity deliberately, or whether a confirmed incident that propagates does it for them.

Based on survey responses from 116 qualified enterprise respondents (100+ employees), drawn from a single July 2026 wave. This is a directional signal from a self-selected sample, not a probability sample. Respondents include managers, individual contributors, VPs/directors, and C-suite leaders, across technology, healthcare, financial services, and other industries.

Across 108 enterprises, trust in automated agent evaluation rose sharply in July — and the failure rate it is supposed to predict did not move at all. The share of organizations that fully trust automated evaluation nearly tripled, from 5% in June to 13%, and the complaint that evaluations don’t match real-world outcomes fell 10 points. Yet the same share as last month — just under half — shipped an agent that passed its evals and then failed a customer. The reason is visible in the cross-tabs: the new trust belongs almost entirely to enterprises that have not yet been burned. Among those that have, 4% trust automated evaluation; among those that haven’t, 24% do. And getting burned does not slow the march to autonomy — it speeds it up.

This is the second wave of the VentureBeat Pulse Research agent reliability tracker, and the first fielded on an instrument identical to the month before it. That makes July the first read on direction rather than position: what moved, what held, and what the movement means.

What moved is confidence. In June, only 5% of enterprises said they fully trusted automated evaluation, and the most-cited limitation was that evaluations align poorly with real-world outcomes (29%). In July, 13% fully trust automated evaluation and the alignment complaint has fallen to 19%, no longer the leading objection. Both shifts are large enough to read as real rather than noise.

What held is the failure. Just under half of organizations (49%) deployed an agent or LLM feature in the past year that passed internal evaluations and then caused a customer-facing failure — statistically indistinguishable from June’s 50% — and a quarter (24%) have seen it happen more than once. Confidence improved; correctness did not. That is the July gap: not between autonomy and trust, as in June, but between trust and the evidence for it.

The cross-tabs explain where the new confidence comes from, and it is not from better evaluations. Trust is concentrated almost entirely among enterprises that have not experienced a false-confidence failure: 24% of them fully trust automated evaluation, against 4% of those that have. The trust curve is being lifted by inexperience. Meanwhile the enterprises that have been burned are not retreating from autonomy — 85% of them already allow zero-human deployment or are engineering toward it, against 61% of those that have not been burned. Overall the autonomy trajectory is flat at 67%, but the population inside it has shifted toward the organizations with the most direct evidence that evaluations miss things.

The vendor market, by contrast, is finally showing signs of settling. The share of enterprises running no dedicated evaluation tooling fell from 17% to 12%; specialist platforms gained, with Braintrust nearly doubling to 15% and DeepEval reaching 17%; and switching intent cooled, with those planning no change rising from 36% to 44%. Selection criteria moved with it: ease of integration overtook cost as the top factor, jumping from 27% to 39%. Enterprises are done shopping on price and have started buying on fit.

Methodology

VentureBeat fielded this survey as part of its ongoing Pulse Research series. This wave — the agentic reliability and evals tracker — examines how technical leaders evaluate agent performance and reliability. Responses are filtered to organizations with 100 or more employees (n=108), drawn from a July 2026 fielding. Because the July instrument is identical to June’s, this report makes month-over-month comparisons where they are warranted; where questions were multiple-select, shares can sum to more than 100%.

Comparisons against June (n=157) are tested for significance, and only a handful of the month’s movements clear a conventional threshold: the rise in full trust in automated evaluation (5% to 13%), the fall in the real-world-alignment complaint (29% to 19%), the jump in ease of integration as a selection factor (27% to 39%), and the gain in Braintrust as a primary platform (8% to 15%). Movements described in this report as flat — the failure rate, the autonomy trajectory, the production monitoring mix, the investment ranking — are statistically indistinguishable between waves, and that stability is itself the finding. Differences of a few points elsewhere should be read as sample variation, not trend.

By role the sample is senior and buyer-credible: 44% are final decision-makers for AI purchases and another 25% recommenders or influencers, a slightly more senior mix than June. Product and program managers (18%), consultants and advisors (12%), CIOs/CTOs/CISOs (11%), and directors of engineering/IT (11%) lead the named titles, alongside a large “Other” function (30%). By organization size the sample is again mid-market-weighted: 100–499 (33%) and 500–2,499 (30%) employees lead, with 2,500–9,999 (23%), 10,000–49,999 (9%), and 50,000+ (5%) above them.

One composition change is worth flagging because it bears on the trust finding. The industry mix shifted between waves: Technology/Software fell from 23% of the June sample to 14% in July, while Retail/Consumer rose from 15% to 19% and now leads. A less technology-weighted sample plausibly carries less hands-on exposure to agent evaluation, and some of the month’s rise in trust may reflect who answered rather than what changed. The burned-versus-unburned split reported in Finding 2 holds within the July sample regardless, but readers should treat the headline trust movement as directional.

At 108 respondents the sample is large enough to support directional conclusions but should not be treated as a precise measurement; it is self-selected and is not a probability sample. Cross-tabs reported here rest on subgroups of 40 to 68 respondents and are correspondingly coarse.

Finding 1: The failure rate did not move

Just under half still ship agents that pass evals and fail customers

We asked whether, in the past 12 months, organizations had deployed an agent or LLM feature that passed their internal evaluations but then caused a customer-facing failure. The answer is the same as last month.

Forty-nine percent of organizations shipped an AI feature that cleared internal evaluations and then failed in front of a customer — an incorrect output, a broken workflow, or a quality incident — against 50% in June. A quarter (24%) have seen it happen more than once, unchanged. Across two waves and 265 enterprises, the rate at which evaluations certify agents that then fail is stable to within a percentage point.

That stability is the anchor for everything that follows. Every other movement this month — rising trust, consolidating tooling, shifting purchase criteria — has to be read against a failure rate that has not responded. Whatever enterprises did between June and July, it did not change how often a passing evaluation turns out to be wrong.

Finding 2: Trust rose — among those who haven’t been burned

Full trust nearly tripled, and the alignment complaint fell ten points

We asked which limitation most reduces trust in automated agent evaluations today. The distribution shifted materially from June.

Two things moved together: Full trust in automated evaluation nearly tripled, from 5% to 13%, and the objection that most directly describes a false-confidence failure — poor alignment with real-world outcomes — fell from 29% to 19%, surrendering the top spot to evaluation bias and inconsistency (22%), now tied with data-leakage concerns (22%). On the surface this reads as an evaluation layer beginning to earn its keep.

The cross-tab says otherwise. Splitting the sample by whether an organization has actually experienced a false-confidence failure, trust divides almost completely. Among the 53 enterprises that shipped an agent which passed evals and then failed a customer, 4% fully trust automated evaluation. Among the 41 that have had no such failure, 24% do — a six-fold difference, and the sharpest split in the dataset. Direct contact with the failure mode is what removes the trust.

This is the month’s central caution. The improvement in sentiment is not evidence that evaluations got better; the failure rate in Finding 1 rules that out. It is what a trust curve looks like when a cohort of less-burned organizations enters the sample and reports its priors. Enterprises reading their own rising confidence as validation of their evaluation stack are reading a number that measures inexperience.

Finding 3: Being burned accelerates autonomy rather than restraining it

85% of the burned are on the zero-human path, against 61% of the REST

We asked whether organizations would let an autonomous agent deploy a code or system change to production on automated evaluation results alone, with no human-in-the-loop validation. The aggregate held; the composition did not.

At the top line, nothing changed: 67% of organizations either already allow zero-human-in-the-loop deployment for low-risk agents (37%) or are actively engineering their pipelines to permit it within a year (30%), against 67% in June. The share ruling it out for the foreseeable future slipped from 22% to 18%. The autonomy ceiling stopped rising, but it did not come down.

Underneath, the picture inverts the intuitive one. Among enterprises that have shipped an evaluation-passing agent that then failed a customer, 85% are on the autonomy trajectory. Among those that have not, 61% are. Organizations with direct, expensive evidence that their evaluations miss things are substantially more likely to be removing the human check, not less — and only 11% of them rule out full automation, against 24% of those that haven't been burned. The pattern is identical for those burned once and those burned repeatedly.

The most plausible mechanism is not recklessness but maturity: the organizations that ship agents at enough volume to hit a customer-facing failure are the same ones with pipelines sophisticated enough to automate, and they are treating the failure as a cost of operating rather than a reason to stop. That is a defensible read. It is also precisely the dynamic that turns Finding 1’s stable failure rate into a growing absolute number of incidents, since the enterprises most likely to fail are the ones scaling their capacity to deploy without review.

One June finding did not replicate. Last month, larger enterprises appeared slightly further down the autonomy path than smaller ones (70% versus 64%). In July the two converge — 65% for organizations with 2,500+ employees against 68% below that, with near-identical failure rates (48% and 50%) — which suggests the June gap was sample variation rather than a size effect. Company size is not what separates the aggressive adopters; experience of failure is.

Finding 4: The stack begins to consolidate

Specialists gain, and the “Nothing at all” share shrinks

We asked which agent reliability or evaluation platform enterprises primarily use today. The field is still crowded, but it is no longer tied at the top with nothing.

The most consequential number is the one that fell. In June, having no dedicated agent-evaluation tooling was tied for the most common answer at 17%; in July it is 12% and fifth. Enterprises are acquiring evaluation tooling, and the specialists are capturing most of that movement: Braintrust nearly doubled its share of primary usage to 15%, and DeepEval reached 17%. Provider-native tooling held roughly flat — OpenAI at 18%, Anthropic at 12% — meaning the growth came at the expense of running nothing rather than at the expense of the model providers.

Counting any use rather than primary platform, the footprints are wider and the ordering is similar: OpenAI native evals reach 31% of enterprises, DeepEval 27%, Braintrust 22%, Anthropic native evals 20%, custom in-house tooling 14%, and Weave and Langfuse 11% each. Nineteen percent still report using no dedicated tooling anywhere in their stack. The category now has three plausible independent contenders where in June it had none with double-digit primary share — the first evidence in this series of an evaluation layer starting to take shape.

Finding 5: Production monitoring still watches the wrong thing

Half monitor whether the agent runs; under a third monitor whether it’s right

Production monitoring for an AI agent can watch two very different things. It can watch whether the system is functioning — is the agent up and responding, did each request complete, how fast, at what cost, with any errors. Or it can watch whether the agent’s output is correct — automated checks that evaluate the content of each answer as it goes out. A confidently wrong answer is invisible to the first kind: the request completes, the response is fast, no error is thrown, and every functioning-metric reads healthy. We asked which kind live production monitoring is built for today.

Grouped by what is actually being watched, the split is essentially June’s: 50% of organizations monitor only whether the agent is functioning, while 26% run automated checks on whether its answers are right. Counting ad-hoc reviewers and don’t-knows, nearly three-quarters of organizations have no automated, real-time evaluation of output correctness in production. Inline quality assertions and transaction trace logging are tied as the most common approach at 26% each on a base of 106 — no single monitoring posture leads.

This is the finding that most directly contradicts the month’s rising confidence. Trust in automated evaluation went up eight points while the runtime capacity to detect an evaluation being wrong went nowhere. Among enterprises that already permit zero-human deployment, only 28% run inline quality checks on production traffic — which means the majority of organizations that have removed the human from the deployment decision have also not replaced that human with anything watching output quality afterward. The gate is automated and the alarm is not installed.

Finding 6: Bought on fit now, not on price

Ease of integration overtakes cost as the top selection factor

We asked what most influenced enterprises’ choice of an evaluation vendor, and what they treat as their primary measure of success. One answer moved sharply; the other did not move at all.

Ease of integration jumped 12 points to 39% and displaced cost as the leading selection criterion, the clearest purchasing shift in the data. Evaluation accuracy rose modestly to 28%, cost fell to 23%, and breadth of observability (6%) and vendor roadmap (2%) remain marginal. Read alongside Finding 4, the two move together: enterprises adopting their first dedicated evaluation tooling are optimizing for what will slot into an existing pipeline this quarter, not for what is cheapest or most capable in the abstract. That is what a market looks like when it stops evaluating and starts installing.

What did not move is what enterprises want from the tool once installed. Evaluation consistency remains the primary success metric at 38%, essentially identical to June’s 36%, well ahead of reduction in failures (20%), speed of experimentation (18%), production visibility (16%), and compliance (7%). The priority is still repeatability — the same verdict on the same behavior every time — which is notable given that bias and inconsistency is now the top-cited trust limitation in Finding 2. Enterprises are buying for integration and measuring for stability, and are not yet getting the second. Satisfaction with current tooling remains moderate, averaging 3.9 on a five-point scale across overall satisfaction, ease of implementation, and value for money, barely changed from June’s 3.8.

Finding 7: Human review becomes the top line item

And the enterprises that have been burned fund it hardest

We asked which reliability and evaluation investment will grow most over the next year. Human review edged into first place.

Human review workflows (31%) and production observability (30%) swapped positions at the top, a change small enough to be noise on its own — but the underlying pattern is the same one June identified and it has strengthened. Enterprises plan to grow spending on human reviewers faster than on the automated evaluation pipelines (19%) that would replace them, at the same moment two-thirds are engineering the human out of the deployment decision. Only 6% report a flat budget, down from 8%.

The cross-tab makes the hedge explicit. Among enterprises that have shipped an evaluation-passing agent that failed a customer, 38% name human review as their fastest-growing investment; among those that have not, 24% do, and they favor observability tooling instead. So the burned cohort is doing both things at once: it is the most aggressive on autonomy (85% on the zero-human path, per Finding 3) and the most committed to funding human reviewers. That is not a contradiction so much as a strategy — automate the deployment decision, and pay people to catch what the automation misses. Whether that scales is the open question, since human review is the one part of the stack that does not get cheaper as agent volume grows.

Finding 8: The switching wave cools

Those planning no change rise from a third to nearly half

We asked whether enterprises plan to adopt a new, additional, or replacement evaluation platform, and which they are considering. Fewer are shopping than last month.

A majority (56%) still intend to adopt a new, additional, or replacement platform within twelve months, but that is down from 64%, and the near-term cohort thinned from 31% to 24%. The share standing pat rose from 36% to 44%. Neither movement clears a significance threshold on its own, but both point the same direction, and they point it consistently with Finding 4: as enterprises actually acquire tooling, the population still looking for it shrinks.

The consideration set has reordered, too. Among the 60 enterprises planning a change, OpenAI’s native evals lead what they are evaluating (20%), followed by Braintrust (18%), Weights & Biases Weave (12%), and DeepEval (10%), with a further 10% actively evaluating but holding no shortlist. DeepEval led June’s consideration set at 20%; it has since converted much of that interest into primary usage, which is what a consideration-to-adoption handoff looks like. Braintrust now occupies the position DeepEval held — high interest ahead of installed base — and is the vendor to watch in the next wave.

The bottom line: Confidence moved, correctness didn’t

June found a gap between the autonomy enterprises were granting their agents and the trust they placed in the evaluations meant to govern it. July finds that gap closing from the wrong side. Trust rose — full confidence in automated evaluation nearly tripled and the complaint that evaluations miss reality fell ten points — while the thing that trust is supposed to track held exactly still. Just under half of enterprises still ship agents that pass their evals and then fail a customer, the same as last month.

The cross-tabs locate the new confidence precisely, and it is not in the evaluations. Twenty-four percent of enterprises that have never had a false-confidence failure fully trust automated evaluation; 4% of those that have do. Trust in this market is a function of exposure, not of evidence. And exposure does not produce caution: the burned cohort is the most autonomous in the sample, with 85% already deploying without human review or building toward it. What it produces instead is a hedge — the same organizations fund human review workflows hardest, at 38%, while removing humans from the deployment gate.

The vendor market is the month’s genuinely encouraging story. Running no dedicated tooling fell from 17% to 12%, specialists gained real share for the first time in this series, buyers shifted from price to integration fit, and switching intent cooled as adoption completed. An evaluation layer is finally forming. But the runtime picture has not followed: half of enterprises still monitor only whether their agents are running, and among those that already deploy without human review, just 28% run real-time checks on output quality.

At 108 respondents in a mid-market-weighted, self-selected sample, and with an industry mix that shifted away from technology between waves, this is a directional read. The direction, though, is legible: enterprises are tooling up, buying for fit, and growing more confident — and none of that has yet changed how often a passing evaluation turns out to be wrong. The question this series carried out of June was whether assurance would catch up to autonomy. July’s answer is that confidence caught up first, which is the harder problem, because an enterprise that trusts a broken gate has less reason to fix it than one that knows the gate is broken.

This report presents the July 2026 wave of an ongoing longitudinal series on enterprise AI agent reliability and evaluation, based on 108 qualified respondents at organizations with 100 or more employees. Comparisons are drawn against the June 2026 wave (n=157), fielded on an identical instrument. At this sample size, results should be read as a directional signal rather than a precise measurement — the sample is self-selected, not a probability sample. Respondents span final decision-makers, technology recommenders/influencers, and end business users, across a mid-market-weighted range of industries and company sizes.

Across 101 enterprises, the context feeding AI agents is failing often and repeatedly. Sixty-eight percent have traced a confident but wrong agent answer to missing or inconsistent business context in the past six months, and the single most common answer is not "once" but "more than once." The counterintuitive part is which companies report it. Enterprises building or running a governed semantic layer (a layer of company-specific definitions and relationships) report recurring failures at more than twice the rate of those without one.

The infrastructure built to fix bad context is, so far, mostly revealing how much bad context there is. Meanwhile the architecture meant to solve the problem commands no consensus at all: hybrid retrieval and outright pluralism finish one respondent apart, in a dead heat.

This wave of VentureBeat Pulse Research examines the enterprise RAG and context layer: what feeds AI agents their business context, which retrieval systems enterprises run, how they buy and measure them, where the architecture is heading, and — most revealingly — how often that context is already failing them.

The central finding is that the context failure is no longer an incident; it is a condition. Sixty-eight percent of enterprises say that in the past six months their AI agents produced confident but wrong answers they traced to missing or inconsistent business context rather than to model error. More striking than the total is its shape: 37% report the failure recurring, against 32% who saw it once. Among enterprises in a position to answer at all, the most prevalent experience of running agents on company data is being wrong repeatedly for reasons that have nothing to do with the model.

The remedy the industry has settled on — a governed semantic or context layer giving agents and BI a shared understanding of the data — is being built at scale: 32% run one in production, another 31% are piloting or building one, and 20% more are evaluating. But the cross-tabs deliver an uncomfortable result: Enterprises that have built or are building a layer report recurring context failures at 50%, against 21% for those without one. The layer isn't causing the failures — it's catching them, which makes it the most useful finding in the wave. The semantic layer is what makes a context defect traceable. Organizations without one are not having fewer failures so much as attributing fewer failures.

Underneath, the stack is unsettled in a way it was not expected to be. Retrieval remains the leading primary context source at 31%, and provider-native retrieval — OpenAI's file search (46%) and Google Vertex AI Search (41%) — still runs well ahead of every dedicated vector database. But the expected architecture has no majority behind it: hybrid retrieval (30%) and "multiple architectures, chosen by use case" (29%) are separated by a single respondent. And enterprises remain firmly unwilling to hand the context layer to a provider — just 12% intend to consolidate onto a single model provider’s native context stack, against 37% holding to best-of-breed and 37% planning an explicit mix.

The buying criteria are where the failure is starting to register commercially. Access control and permissions is now tied with ease of data ingestion as the top selection factor at 24% each, and response correctness is the primary success metric for 38% of enterprises. Enterprises are beginning to buy retrieval for the properties that govern context rather than the properties that move it.

Methodology

VentureBeat fielded this survey as part of its ongoing Pulse Research series. This survey focused on enterprise RAG infrastructure and the context layer — the retrieval systems, semantic layers, and context sources that feed AI agents. Responses are filtered to organizations with more than 100 employees (n=101). All responses are from a single July 2026 wave, so the report reads cross-sectionally and does not infer month-over-month trends. Several questions were multiple-select; those shares are reported as a percentage of respondents, not of total selections, so they can sum to more than 100%.

By organization size the sample concentrates in the mid-market: 101–250 employees (34%), 1,001–5,000 (25%), and 251–1,000 (25%) lead, with 10,001+ (12%) and 5,001–10,000 (5%) above them. By role it spans managers (39%), individual contributors (29%), VPs and directors (22%), and the C-suite (9%); on purchasing authority it is buyer-credible, with 38% final decision-makers and another 43% recommenders or influencers. Technology/Software is the largest industry at 31%, followed by Healthcare/Life Sciences (14%), Retail/E-commerce (10%), and Manufacturing (9%).

A note on the context-failure base: Of the 101 respondents, 10 either do not run agents on enterprise data (5%) or do not trace root cause at that level (5%). Headline shares for the failure question are reported on the full 101; the subgroup comparisons in Finding 2 use the 91 respondents who were able to give a yes-or-no answer, since including those who cannot observe the failure would bias the comparison toward whichever group is less instrumented. Subgroup cells run from roughly 10 to 62 respondents and are correspondingly coarse; where a cell falls below 10 it is not reported as a percentage. A small number of respondents selected "Other" and gave a write-in industry (6%) or role (2%) that didn't map to a listed category; those shares appear as not stated in the appendix rather than being redistributed.

At 101 respondents this is a modest sample and should be read as a directional signal rather than a precise measurement; it is self-selected and is not a probability sample. It is best read as the view from organizations actively standing up RAG and context infrastructure rather than from the largest operators.

Finding 1: Confident, wrong, and repeating

The most common answer isn't "once" but "more than once"

We asked whether, in the past six months, enterprises had traced a confident but wrong agent answer to missing or inconsistent business context rather than to model error. Most had — and most of those had seen it happen again.

This is the report’s defining number. Sixty-eight percent of enterprises have had an AI agent produce a confident, wrong answer they traced to bad context — wrong metric definitions, stale data, missing documents — and the recurring case (37%) outweighs the one-off (32%). Only 22% report no such failure. Restricted to the 91 enterprises able to observe and attribute the failure at all, 76% have experienced it and 41% repeatedly.

The failure mode is specific and dangerous precisely because it does not look like a failure. The model is not visibly hallucinating; it is confidently wrong because the context feeding it was thin, stale, or inconsistent — and it delivers that wrong answer with the same authority as a right one. That the modal experience is recurrence rather than a single incident matters more than the headline share: a one-time failure is an incident to be fixed, while a repeating one indicates a structural defect in how business context reaches the agent. Everything else in this report — what enterprises retrieve, how they govern it, and what they plan to build — is downstream of this problem.

Finding 2: The semantic layer reveals the failure before it fixes it

Enterprises building a governed layer report more recurring failures, not fewer

We asked whether enterprises use a governed semantic or context layer to give agents and BI a shared understanding of their data. Most are on the path — and cross-tabbing that answer against the failure in Finding 1 produces the wave’s most counterintuitive result.

Engagement with the governed context layer is broad. Sixty-three percent of enterprises either run one in production (32%) or are piloting and building one (31%), and a further 20% are actively evaluating, meaning more than four in five are engaged with the idea in some form. Only 14% have no plans.

The cross-tab is where it gets interesting. Among the 91 enterprises able to answer the failure question, those who have built or are building a semantic layer report recurring context failures at 50%, while those without one — evaluating or with no plans — report them at 21%, a gap that clears conventional significance thresholds (p=0.01) and runs in the direction opposite to what the technology is sold to do. Narrowing to enterprises with a layer specifically in production points the same way but does not carry statistical weight on this sample: 53% recurrence against 34% for everyone else, a difference that does not reach significance and should be read as directional only.

Read as causation, this is implausible — a governed definition layer does not manufacture wrong answers. Read as detection, it is the most useful result in this wave. Tracing a confident wrong answer to a specific context defect — a metric defined two ways, a stale table, a document the agent could not see — requires exactly the shared, governed definitions a semantic layer provides. Without one, the same failure occurs and gets logged as a model problem, a user error, or nothing at all. The causation almost certainly also runs backwards in part: enterprises that have been burned repeatedly are the ones who went and built the layer.

The size split points the same way, and carries significance where the production split does not. Enterprises above 1,000 employees report recurring context failures at 55%, against 30% of those between 101 and 1,000 (p=0.02) — despite the larger organizations being less likely, not more, to have a semantic layer in production (24% against 37%). Larger enterprises have more instrumentation, more auditing, and more people whose job is to ask why a number was wrong. The practical implication for readers is uncomfortable but clear: a low reported context-failure rate is not evidence of a healthy context layer. It is at least as likely to be evidence that nobody is looking.

Finding 3: RAG leads as the context source — and carries the failures

Retrieval feeds more agents than anything else, and fails a large share of them

We asked what an enterprise’s AI agents primarily use to understand its data. Retrieval leads, but no longer by the margin the category assumes.

 Retrieval remains the backbone of enterprise context at 31%, ahead of a governed semantic layer (19%) and mixed approaches (17%). But the tail has thickened in a way worth noting: long-context loading is now the primary source for 13% of enterprises, and 5% let agents run on the model’s general knowledge with no enterprise context layer at all. Between them, nearly one in five enterprises is feeding agents business context either by brute-force context window or not at all.

Cross-tabbed against Finding 1, the sources do not fail equally. Among enterprises whose primary context source is retrieval, 87% report a context-traced failure and 48% report it recurring — on the largest base of any group, 31 respondents. Those relying on a governed semantic layer report 79% and 53%; mixed approaches 79% and 36%; direct live-system queries 40% and 30%. The long-context group is the outlier in the other direction, reporting 64% any failure but only 9% recurrence.

These subgroup figures should be read with the detection caveat from Finding 2 firmly attached. Groups differ in how well they can attribute a wrong answer to a context defect as much as in how often they suffer one, and the cells here run from 10 to 31 respondents. The retrieval group’s 87% is best read as evidence that RAG-heavy enterprises both experience and notice context failures, not as a clean measurement of relative reliability.

What survives the caveat is the structural point. Because so much enterprise context flows through retrieval, and because retrieval carries that load on the widest base in the sample, the quality of retrieval is the quality of the answer. When RAG is the default source, incomplete retrieval is the main point of failure.

Finding 4: Model-backed and hyperscaler retrieval still leads the vector databases

OpenAI's file search and Google's Vertex AI Search top every purpose-built system

We asked which retrieval systems enterprises run in production today. The answer continues to favor the model providers and hyperscalers over the specialists.

The dedicated vector database is not the center of the RAG stack. OpenAI’s file search (46%) and Google’s Vertex AI Search (41%) lead by better than three to one over any purpose-built alternative. Among the specialists, the most-used remain the ones enterprises already run for other reasons — Elasticsearch/OpenSearch at 20% and pgvector at 15% — while the pure-play vector databases that define the category (Pinecone, Weaviate, Milvus, Qdrant) each sit between 7% and 12%. Custom in-house retrieval stacks, at 18%, outrank every pure-play vendor.

Which system is actually primary separates retrieval from infrastructure

Usage counts alone understate the gap, because enterprises run several of these systems at once. We also asked which one is primary. The share of each system’s own users who name it their primary retrieval platform divides the field cleanly.

Elasticsearch and pgvector are widely present and rarely primary: four in five of their users retrieve mainly through something else. They are infrastructure the enterprise already ran, pressed into service at the edges of a retrieval stack whose center is elsewhere.

Model-backed and hyperscaler retrieval is not merely the most common system on the list; for most of the enterprises that adopt it, it is the system of record. Custom in-house stacks behave the same way — when an enterprise builds one, it is usually the primary, not a side project.

The primary-platform question was fielded as a single-select and 18 of 101 respondents selected more than one option, so the shares above are computed as a proportion of each system’s users rather than of the full sample. On the 83 respondents who gave exactly one answer, the ranking is unchanged: OpenAI's file search 28%, Vertex AI Search 23%, custom in-house stack 12%, and no pure-play vector database above 8%.

The comparison worth sitting with is what this leaves for the RAG specialists. In a category built around specialist infrastructure, more enterprises have written their own retrieval stack than run any single dedicated vector database — and roughly four times as many use retrieval that arrived bundled with a model provider or cloud they already buy from. Only 7% run no production RAG at all, so this is not a story about early adoption; it is a story about where retrieval gets acquired.

Finding 5: No architecture commands a consensus

Hybrid retrieval and "It depends on the use case" finish in a dead heat

We asked which retrieval architecture enterprises expect to dominate their production RAG systems by the end of 2026. No single answer comes close to a majority — and the two front-runners are separated by one respondent.

Hybrid retrieval leads at 30%, with the expectation that no single architecture will dominate at all immediately behind at 29%. The gap is one respondent, far inside the margin on a sample this size, and the honest reading is that these two finish level rather than that either is in front. Together they account for 58% of enterprises, and what unites them is more instructive than what separates them — both describe layered pipelines rather than a single retrieval technique, and neither expects the pure vector-search approach that launched the category to carry production on its own.

Two smaller answers carry the sharper signal. Fifteen percent expect tool-first or long-context retrieval to dominate without a dedicated vector layer at all — a direct challenge to the premise of the category — while 12% still expect vector-only retrieval to prevail. That the anti-vector position now edges the pure-vector one, on a three-respondent margin that is itself too narrow to call, is a notable inversion for an industry that spent three years building vector databases. Add the 15% who are unsure or expect no large-scale RAG, and the picture is of a market that agrees vector search alone is insufficient and has not agreed on what replaces it.

Finding 6: Enterprises decline to hand the layer to a provider

Consolidation onto a provider's native context stack barely registers

We asked how enterprises will respond as model providers bundle retrieval, memory, and orchestration into their platforms. Their stated intent cuts sharply against their current usage.

Here is the tension at the heart of the stack. Provider-native retrieval leads actual usage by a wide margin (Finding 4), yet just 12% of enterprises intend to consolidate onto a provider’s native context stack. Best-of-breed standalone tools and an explicit mix are tied at the top at 37% each, and 6% intend to build and own the layer themselves — meaning 79% of enterprises expect to keep at least part of the context layer outside any single provider.

The gap between what enterprises run and what they say they want is the strategic question of the category. They are adopting bundled retrieval because it arrives with tools they already buy, while asserting they will preserve independence. Read against Finding 2, the stated preference has a rationale beyond vendor politics: the failures enterprises are trying to fix are failures of governed, consistent, access-aware business context, and that is precisely the layer they are least willing to outsource. Whether the preference survives contact with the convenience of the bundle is what the next several waves will decide.

Finding 7: Access control climbs into the buying decision

Governance now ties ingestion as the reason a system gets chosen

We asked what matters most when enterprises choose a retrieval system, and what they treat as the primary measure of success once it is running.

The selection criteria have moved toward governance. Access control and permissions (24%) is now exactly tied with ease of data ingestion (24%) at the top, ahead of retrieval accuracy and latency and performance (15% each) and operational simplicity (14%). That puts a governance property at the top of the purchase decision for the first time in this series — and it is the property most directly implicated in the confident-but-wrong failures of Finding 1, where an agent surfaces something it should not have seen or misses something it should have.

Once systems are running, the emphasis on correctness is unambiguous: response correctness is the primary success metric for 38% of enterprises, twice the next answer, security and access control (19%). Answer relevance (17%), latency (13%), and operational stability (11%) trail. Taken together, 56% of enterprises measure their retrieval system primarily on whether its answers are right or properly permissioned, rather than on whether it is fast or stable.

Satisfaction with current systems is moderately positive: on a five-point scale, overall satisfaction averages 4.13, value for money 4.01, and ease of implementation 3.98. That is a respectable set of scores for a layer that, on this wave’s evidence, is producing recurring wrong answers in nearly four in ten enterprises — which suggests enterprises are rating the tools against expectations of what retrieval infrastructure does, not against the outcome of getting the answer right.

Finding 8: Half the market is in motion

Vertex AI Search leads the consideration set — and so does uncertainty

We asked whether enterprises plan to change or add a retrieval provider, and which they are considering. The consideration set is broader than today’s stack.

The retrieval stack is not settled, but it is not churning, either: about half of enterprises have no plans to change, while the other half — 52 of 101 — intend to switch or add a provider within twelve months, a fifth of them within the next quarter. Among those 52 enterprises in motion, Google’s Vertex AI Search leads the consideration set at 35%, followed by Elasticsearch/OpenSearch (25%), Pinecone (23%), and OpenAI's file search (23%).

Two patterns stand out. First, the pure-play vector specialists draw markedly more forward interest than their current footprint would suggest — Pinecone is considered by 23% of movers against 12% present usage, Weaviate 17% against 10%, Qdrant 15% against 7%, and Milvus 14% against 9%. The specialists are not winning the installed base, but they are firmly in the evaluation, and each of them roughly doubles its footprint in forward consideration. Second, 15% of movers are evaluating with no shortlist at all and 17% are considering a custom in-house stack — together nearly a third of enterprises planning a change either do not know what they want or intend to build it.

The bottom line: A context failure that better detection is only beginning to reveal

Organizations with more than 100 employees are running agents on business context they cannot yet guarantee, and the evidence has moved past anecdote. Sixty-eight percent have traced a confident, wrong agent answer to missing or inconsistent context in the past six months, and the recurring case now outweighs the one-off. Retrieval remains the default source of that context and carries the failure on the widest base in the sample — while nearly one in five enterprises has fallen back to long-context loading or the model’s general knowledge, which is not a context layer at all.

The most important result in this wave is the one that inverts the expected direction. Enterprises building or running a governed semantic layer report recurring context failures at 50%, against 21% for those without one, and larger enterprises report them at nearly twice the rate of mid-market peers despite being less likely to have the layer built. The straightforward reading is that instrumentation reveals failures rather than causing them, and that the organizations reporting clean context records are largely the ones without the means to check. That reframes the entire finding: the 22% reporting no context failure are not the well-governed cohort, and a low failure rate should be treated as a question rather than an answer.

Meanwhile, the fix has not converged. Hybrid retrieval and architectural pluralism finish level as the expectation for production RAG by the end of 2026, one respondent apart; the anti-vector position narrowly edges the pure-vector one; and while provider-native retrieval leads usage by a wide margin — and is the primary system for most of the enterprises that run it — only 12% will consolidate onto a provider’s stack, with 79% keeping some part of the layer independent. The commercial signal is that access control has climbed to tie ease of ingestion as the top buying criterion, and response correctness is the dominant success metric — enterprises are starting to buy retrieval for the properties that govern context rather than the ones that move it.

At 101 respondents in a single July wave, skewed toward the mid-market, this is a directional read. But the direction is clear enough to act on: the context layer is the contested tier of the AI stack, the failure it produces is recurring rather than occasional, and the enterprises best equipped to see the problem are the ones reporting it worst. The open question for later waves is whether the governed context layer starts to reduce the failures it is currently so good at exposing.

Based on survey responses from 101 qualified enterprise respondents (100+ employees), drawn from a single July 2026 wave. At this sample size the results should be read as a directional signal rather than a precise measurement — this is a self-selected sample, not a probability sample. Respondents include managers, individual contributors, VPs/directors, and C-suite leaders.

Across 107 enterprises, agentic orchestration is not a choice of a single platform.

The typical enterprise runs three orchestration platforms at once, and selects them for flexibility across models rather than affinity to any single one. Microsoft leads primary usage while Anthropic leads forward consideration by a wide margin. 

The AI control plane enterprises expect is deliberately hybrid, meaning it includes use of the leading AI providers, but also provider-independent technologies — and the risk they fear most from provider-resident control is not lock-in but the provider’s own security and permissioning limits. 

One in five enterprises still has no real-time way to stop a runaway agent before the bill arrives.

This wave of VentureBeat Pulse Research examines enterprise agent orchestration: which platforms enterprises run on, what drives the choice, what they optimize for, how they expect agent control to be structured, and — most revealingly — how orchestrated their deployed “agents” actually are and how tightly they control the cost of running them.

The central finding is that orchestration has become plural. Eighty-five percent of enterprises run two or more orchestration platforms and 64% run three or more, with a mean of 3.1 platforms per organization. Microsoft AI Foundry / Copilot Studio appears in 70% of stacks and OpenAI’s Agents SDK in 68%, with Anthropic’s Claude Platform in 47%. Asked to name a single primary platform, respondents who gave one unambiguous answer put Microsoft first (41%) and Anthropic second (28%). Nobody in this sample is running one orchestration layer and calling it a strategy.

The selection logic follows from that plurality. Flexibility across models and tools is the leading purchase driver at 29%, nearly three times the share naming model gravity — native alignment with a state-of-the-art base model — at 10%. Enterprises are not choosing the orchestration environment that comes with their favorite model; they are choosing the one that does not commit them to any model. Security and permissions (17%), production reliability (15%), and control over agent execution (15%) fill out a buying logic focused on governance and optionality rather than developer convenience.

A clear majority (53%) expect a hybrid control plane by the end of 2026 — provider-native plus external orchestration — and the risk they most associate with provider-resident control is security and permissioning limitations (37%), ahead of vendor lock-in (23%) and limited visibility (22%). Investment has moved accordingly: agent monitoring and debugging leads the spend at 31%, with security and permissions enforcement at 30%, while workflow tooling draws 19%. Enterprises are spending to see and govern agents, not merely to build them.

Most companies admit that a majority of their “agents” are really just chatbots. A plurality of 47% of respondents say that between 26 and 50% of their agents are genuinely orchestrated, with 37% at a quarter or below and 16% past the halfway mark. 

But fiscal control remains the soft spot: 21% of enterprises track agent spend only through post-hoc logs, with no real-time way to halt a runaway execution loop.

Methodology

VentureBeat fielded this survey as part of its ongoing Pulse Research series, with this instrument focused on enterprise agent orchestration. Responses are filtered to organizations with 100 or more employees (n=107), drawn from a single July 2026 wave; because this is one wave rather than a pooled multi-month sample, the report reads cross-sectionally and does not infer month-over-month trends. All figures in this report come from the July fielding only. Where questions were multiple-select, shares can sum to more than 100%.

This wave draws a notably large-enterprise, technology-heavy sample, and that shapes every finding in it. By organization size, more than half sit at 10,000 employees or above: 50,000+ (26%) and 10,000–49,999 (25%) lead, followed by 2,500–9,999 and 500–2,499 (19% each) and 100–499 (11%). Technology/Software accounts for 53% of respondents, with Government/Public Sector (16%) and Manufacturing/Industrial (10%) next. By role the sample is hands-on and technical: software and ML engineers (22%), product and program managers (21%), directors of data/AI/analytics (17%), and VPs of data/AI/analytics (12%). On purchasing, 90% are recommenders, influencers, or final decision-makers for AI solutions (63% recommender/influencer, 27% final decision-maker).

A note on the primary-platform question. Forty-six of 107 respondents registered more than one selection on a question intended to capture a single primary platform. Because those responses cannot be resolved to one answer, primary-platform shares are reported on the 61 respondents who gave a single unambiguous answer, and are labeled as such wherever they appear. Platform footprint figures — which platforms an enterprise uses at all — use the full n=107 base and are unaffected. The ambiguity is worth noting on its own terms: on a question asking for one platform, more than four in 10 respondents could not or would not narrow to one, which is consistent with the multi-platform pattern documented in Finding 1.

At 107 respondents the sample is robust enough to read directionally with reasonable confidence, though it remains self-selected and is not a probability sample. Because each subgroup here only includes about 50 to 60 respondents, splits between them are less precise than the full-sample findings.

Finding 1: Orchestration is a portfolio, not a platform

The typical enterprise runs three orchestration platforms at once

We asked which agent orchestration platforms enterprises use, and which one they treat as primary. The first answer is that almost nobody has just one.

The defining feature of this layer is plurality. Only 15% of enterprises run fewer than two orchestration platforms; the median organization runs three, and one in six runs five or more. Read that way, the platform “shares” below describe overlapping deployments rather than a divided market — Microsoft and OpenAI each appear in roughly seven of ten stacks precisely because most stacks have room for several.

Asked to name one primary platform, the 61 respondents who gave a single unambiguous answer put Microsoft AI Foundry / Copilot Studio first at 41%, Anthropic’s Claude Platform second at 28%, LangChain / LangGraph at 10%, and OpenAI’s Agents SDK at 7%, with Google, Amazon, Salesforce, and custom in-house builds at 3% each. Microsoft’s lead on primary usage alongside OpenAI’s near-equal footprint on any usage is the signature of an enterprise-weighted sample: the Microsoft platform arrives through an existing enterprise agreement and becomes the default seat of record, while other platforms are added around it for specific work.

A note on reading these shares: As described in the methodology section, the respondents are self-selected, this wave skews heavily toward large technology organizations, and the primary-platform figures rest on a 61-respondent subset. The numbers measure where this cohort has placed its orchestration bets today, within a self-selected audience of AI-active technical practitioners. A sample built this way can diverge substantially from spend-weighted market measures, and each VB Pulse survey draws its own sample with its own company-size and industry mix, so vendor figures should not be compared across our surveys, either.

Respondents rate the platforms they run at 4.17 out of 5 for overall satisfaction, 3.91 for ease of implementation, and 3.63 for value for money — with value for money the weakest of the three by a clear margin. That ordering is itself a finding: enterprises are broadly happy with what these platforms do and distinctly less happy with what they cost, which is the same nerve the fiscal-control finding touches at the end of this report. Satisfaction sits alongside a two-thirds intent to change platforms within the year; this remains a layer enterprises work with rather than settle on.

Finding 2: Flexibility, not model gravity, drives selection

Enterprises buy the orchestration layer that doesn't commit them

We asked what most influenced the orchestration platform choice, and optionality leads by a distance.

Flexibility across models and tools (29%) is the selection-side explanation for the multi-platform reality in Finding 1: enterprises are choosing orchestration environments on the strength of what they leave open rather than what they lock in. Model gravity — picking the orchestration layer that comes with a preferred frontier model — draws just 10%, less than a third of the flexibility share, which places the pull of any single base model well down the list of what actually decides this purchase.

The next tier reinforces the governance emphasis. Security and permissions (17%), production reliability (15%), and control over agent execution (15%) together account for 47% of responses: nearly half of enterprises pick their orchestration platform on whether they can constrain and depend on what it runs. Ease of development draws 8% and total cost of ownership 4%, an inversion of how these platforms are usually discussed in engineering circles. Performance sits last at 2% — at this stage of adoption the binding constraints are optionality and control, not raw speed.

Finding 3: The job is reliable multi-step execution

Enterprises judge orchestration by whether it completes the work

We asked what enterprises optimize for — their primary success metric for orchestration. Reliability and multi-step workflow management lead, with developer productivity closer behind than in the buying criteria.

Task completion reliability (30%) and multi-step workflow management (27%) together account for 57% of responses: orchestration succeeds, in the enterprise view, when it reliably carries a task through multiple steps to completion. Developer productivity takes a substantial 23% — notably higher than ease of development’s 8% as a purchase driver in Finding 2, which suggests enterprises do not expect to buy developer velocity so much as to earn it once the platform is in place. End-user experience is a minor concern at 7%, consistent with orchestration being an internal execution problem rather than a UX one.

This reliability-first standard is the yardstick against which the portfolio-maturity finding later in this report should be read: enterprises define success as dependable multi-step execution, and a little over a third of them still say a quarter or fewer of their deployed agents do multi-step work at all.

Finding 4: Two-thirds plan to move — and Anthropic leads the consideration set

The installed base and the pipeline point to different vendors

We asked whether enterprises plan to adopt a new, additional, or replacement orchestration platform in the next 12 months, and which platforms they are considering.

Two-thirds of enterprises (67%) intend to adopt a new, additional, or replacement orchestration platform within the year, but the clock runs longer than the intent suggests: the largest cohort sits at 6–12 months (28%) and only 15% expect to move within a quarter. This is deliberate re-platforming on a planning horizon, not urgent churn.

The consideration set is where this finding earns its headline. Among the 72 enterprises in motion, Anthropic leads at 43% — well ahead of Google (31%), custom in-house builds (31%), OpenAI (25%), LangChain / LangGraph (17%), and Microsoft (17%). Set that against Finding 1, where Microsoft leads primary usage and appears in 70% of stacks: the installed base and the forward pipeline point at different vendors. Anthropic draws roughly two and a half times Microsoft’s forward consideration despite trailing it on current primary usage, and custom in-house control planes draw as much interest as any external platform besides Anthropic. A further 18% of movers are evaluating with no shortlist at all.

Read alongside the flexibility-first selection logic in Finding 2, the shape of the next twelve months is legible: enterprises expect to add rather than replace, they are shopping for platforms that preserve model choice, and a substantial minority intend to solve the problem themselves rather than buy it.

Finding 5: Investment flows to watching and governing agents

Monitoring and permissions lead the spend; workflow tooling trails

We asked which orchestration-related investment will grow most next year. Observability and governance take the top two places.

Monitoring and debugging (31%) and security and permissions enforcement (30%) are effectively tied at the top and together account for 61% of planned growth. The money is going to seeing what agents do and constraining what they are allowed to do — the two capabilities that matter once agents are running in production rather than being built toward it. Workflow tooling (19%) and scaling infrastructure (18%) trail, and almost no one is standing still: just 3% report a flat budget.

The emphasis is consistent with the buying logic in Finding 2, where security and permissions was the second-ranked selection factor, and with the control-plane architecture in Finding 6. Enterprises that have decided to run agents across three platforms have a visibility and permissioning problem by construction, and they are funding it directly.

Finding 6: The control plane will be hybrid — and security is why

Enterprises split control, and fear the provider's permissioning more than lock-in

We asked where enterprises expect the primary control plane for agents to live by the end of 2026, and what worries them most if that control sits inside a model-provider platform.

Hybrid control is the dominant expectation by a wide margin (53%). Taken together, the hybrid, custom in-house, and externally-abstracted options — every architecture that keeps control at least partly outside the provider — sum to 78% of enterprises, against 14% willing to hand control to a provider-managed service outright.

The reason enterprises give is worth separating from the one usually assumed. Security and permissioning limitations lead the risk question at 37%, well ahead of vendor lock-in at 23%, with limited visibility and observability close behind at 22%. Combining the security and visibility answers, 59% of enterprises name a control-and-oversight concern rather than a commercial one. The worry is less that a provider platform will be hard to leave than that it will not let them see or constrain what their agents are doing while they are on it — the same concern funding the monitoring and permissions spend in Finding 5. Only 2% say provider-resident control is not a concern at all.

Finding 7: The chatbot trap is loosening, not broken

“Bridging the gap” is now the modal answer on portfolio maturity

We asked enterprises to assess their portfolios honestly: What share of their deployed “agents” are true multi-step orchestrated workflows versus simple single-prompt chatbot wrappers.

The center of gravity has moved into the middle band. Just under half of enterprises (47%) now put between a quarter and half of their portfolio in genuinely orchestrated, stateful workflows, and 16% are past the halfway mark. The bottom two bands — a quarter or fewer genuinely orchestrated — account for 37%, and outright pure-chatbot portfolios have nearly vanished at 3%. Against the reliability-first success standard in Finding 3, this is a portfolio that has started to do the work the orchestration layer exists for, without most of it being there yet.

Maturity tracks platform count. Enterprises reporting a quarter or less genuine orchestration run 2.8 platforms on average; those in the 26–50% band run 3.5. The organizations furthest into real multi-step work are the ones running the most orchestration platforms at once, which is the practical case for the flexibility-first selection logic in Finding 2 — multi-step portfolios appear to accumulate platforms rather than converge on one.

One split that might be expected does not appear. Organization size makes no difference to portfolio maturity in this wave: 38% of enterprises at 10,000+ employees report a quarter or less genuine orchestration, against 37% of smaller ones, and the shares past the halfway mark are equally close (16% and 15%). Whatever separates the mature portfolios from the immature ones here, it is not headcount.

Finding 8: Fiscal control is still reactive for one in five

A fifth of enterprises learn about a runaway agent from the logs

Finally, we asked how enterprises enforce fiscal control over agent token consumption — the risk that an autonomous loop exhausts a budget before anyone intervenes. The approaches split four ways, fairly evenly.

One in five enterprises (21%) has no real-time, programmatic way to stop an agent before a budget-breaking bill arrives — they learn of it from the logs afterward. Another 30% lean entirely on the native caps and throttles built into their primary platform, a control only as good as the provider’s tooling and one that sits awkwardly beside the hybrid, keep-control-outside posture of Finding 6. Roughly half of enterprises — those building custom gateways (25%) or exploiting cross-model routing to arbitrage cost (24%) — are treating token burn as an engineering problem to be controlled deterministically, and the routing group is doing so in a way that only works because they run several platforms at once.

Unlike previous waves, no size split appears here: 18% of enterprises at 10,000+ employees exercise only reactive control against 23% of smaller ones, a difference well within sample noise. The gap in fiscal control in this wave is not between large and small enterprises but between those that have built a cost-control plane and those still relying on whatever their provider ships. Read against the satisfaction scores in Finding 1 — where value for money was the weakest of three ratings at 3.63 — the picture is of a cohort that is unhappy about what agents cost and, in half of cases, not yet instrumented to do much about it.

The bottom line: Plural by design, governed by intention, metered by hope

Organizations with 100 or more employees describe an orchestration strategy built around optionality rather than commitment. They run three platforms on average, choose them for flexibility across models rather than affinity to any one, and judge them on whether they carry multi-step work reliably to completion. Microsoft anchors the installed base and appears in seven of ten stacks; Anthropic leads forward consideration by a wide margin among the two-thirds planning a change; and a substantial minority intend to build their own control plane rather than buy one. Today’s footprint describes where these enterprises are, and clearly does not describe where they intend to stay.

The governance posture is deliberate and consistent. A hybrid control plane is the majority expectation, 78% intend to keep control at least partly outside the provider, and the reason is not commercial but operational — security and permissioning limits (37%) and limited visibility (22%) outrank vendor lock-in (23%) as the fear attached to provider-resident control. The budget follows the fear: monitoring and debugging and security and permissions enforcement together take 61% of planned investment growth, ahead of the tooling used to build agents in the first place.

Where the strategy thins out is cost. Portfolio maturity has moved into the middle — 47% now report between a quarter and half of their agents genuinely orchestrated, and pure-chatbot portfolios have nearly disappeared — but 21% still cannot stop a runaway agent in real time, another 30% depend on whatever caps their provider ships, and value for money is the lowest-rated attribute of the platforms they run. Enterprises have worked out how they want agents governed well before they have worked out how to meter them.

At 107 respondents in a single July wave, skewed toward large technology organizations, this reads as a clear directional signal rather than a precise measurement. The questions for subsequent waves are whether the middle band of portfolio maturity keeps climbing, whether the forward consideration for Anthropic and for in-house control planes converts into deployment, and whether fiscal control catches up to a cost that enterprises already say they are not getting their money’s worth on.

Based on survey responses from 107 qualified enterprise respondents (100+ employees), drawn from a single July 2026 wave. This is a self-selected sample rather than a probability sample, and figures should be read directionally rather than as precise measurement. Respondents include software/ML engineers, product/program managers, directors and VPs of data/AI/analytics, enterprise architects, and directors of engineering/IT, across technology/software, government/public sector, manufacturing/industrial, and financial services organizations.

SpaceXAI, the division of SpaceX formerly known as xAI, is launching an early beta version of Grok Bot, a new agent designed to move AI assistants beyond answering prompts and toward continuously executing work across the software employees already use.

The central idea is straightforward: instead of opening an AI assistant whenever a task arises, users create persistent Bots with specific jobs, give them access to applications and websites, and delegate work much as they would to a teammate.

Each Bot operates through its own computer environment, can continue working when the user's laptop is closed, and can return when it needs approval or has finished the assignment.

SpaceXAI says the system began as an internal prototype before spreading across the company, where teams created Bots for sales outbound, marketing campaigns, office operations, bug fixes and other work. The company is now turning that internally developed workflow into a product for external users.

“Bots are AI teammates that do real work for you,” the company said in announcing the product. “They sign in to your tools, use them just like you do, and come back with finished work.”

The company did not release benchmarks for Grok Bot's performance on agentic tasks. And it arrives amid an increasingly crowded marketplace of first-party AI agents that attempt to reliably complete real, enterprise workflows by interfacing with a user's other applications and devices.

Anthropic introduced computer use for Claude in 2024, allowing models to inspect screens and operate interfaces through mouse and keyboard actions, and continued expanding with the launch of the developer focused Claude Code harness in early 2025 and the more non-technical, white collar focused Claude Cowork agent early this year.

Meanwhile, OpenAI gave its Codex harness the ability to control other computer apps in April, launched agentic Workspace Agents that can also connect to third-party applications and use them autonomously, and recently debuted a new ChatGPT Work environment for longer, multi-step tasks and finished deliverables.

Grok Bot seeks to join the party with its own management model for agents: persistent workers with responsibilities, memory, learned routines and the ability to hand work to one another.

Pricing and availability: Grok Bot starts at $120 per seat per month for teams, $200 per month for individuals

Grok Bot is available beginning today, August 11 in beta for SuperGrok Heavy, Cursor Ultra and Cursor Premium Teams subscribers (recall SpaceX acquired Cursor for $60 billion back in June). The product arrives for macOS, Windows, Linux and iOS, with Android listed as coming soon.

According to its product page on xAI.com, Grok Bot is included with Cursor Ultra at $200 per month for individuals. The plan includes a computer for Grok Bot, access to users' tools, scheduled routines, desktop and mobile operation, and extended AI-token limits.

For organizations, Cursor Premium Teams costs $120 per seat per month and adds centralized billing and settings, a team marketplace for skills and plugins, shared usage analytics and SAML/OIDC single sign-on.

Existing SuperGrok Heavy ($300 per month) subscribers also receive access. However, for organizations wishing to sign up today, SpaceXAI is directing them to a waitlist for future access.

Those prices make Grok Bot a substantially different purchasing decision from a low-cost general AI subscription. The economic question for companies will be whether persistent Bots can replace enough manual work or conventional automation infrastructure to justify the per-user cost — and how usage limits affect total cost once agents begin running continuously.

From prompting an AI to managing one

SpaceXAI describes Grok Bot as a team of “always-on agents.” Users can create multiple Bots, assign each a role and let them work simultaneously.

The company provides examples including Sales Outbound, Talent Scout, Paid Media, Expense Manager, Product Performance, Bug Reproduction, Account Health and Chief of Staff. A sales Bot, for example, can research accounts, score prospective contacts, prepare email and LinkedIn outreach in the user's voice, and assemble the results for human approval.

Promotional materials show SpaceXAI using the system internally for substantially longer chains of work. One sales Bot can add call-transcript notes to a CRM and draft follow-up messages. An operations Bot can seat new hires and process invoices arriving through Gmail. An engineering Bot can reproduce a bug in the product interface, file a ticket and then hand the repair to a debugging Bot.

The architecture could make Grok Bot particularly relevant for workflows that span systems that were never designed for AI automation.

Rather than requiring every application to expose an API specifically for an agent, Grok Bot can sign into applications and websites and operate their interfaces. SpaceXAI says Bots have their own computers and can continue working 24/7.

The company explicitly says this includes websites and applications that have “no clean API or MCP,” an important distinction for enterprises with legacy software, fragmented SaaS environments or internal systems that have never been instrumented for agent access. Instead of limiting automation to formally integrated services, Grok Bot is designed to work through the same software interfaces a human employee would use.

The company says early users are already applying Bots to jobs including vendor negotiations, e-commerce customer support and continuously updating CRM systems.

Another feature attempts to reduce the engineering required to automate repeatable business processes. Users can demonstrate a workflow while a Bot follows along. Grok Bot can then save the process as a routine and execute it later without requiring the user to reproduce every instruction.

SpaceXAI says the Bot can also incorporate corrections into those learned routines, allowing the workflow to change as the user teaches it how a particular process should be handled.

That potentially changes the deployment model from explicitly programming an automation to teaching an agent how an employee performs the job.

The company is also claiming a more persistent form of behavioral memory than simply retaining a chat transcript. According to the launch announcement, Bots remember prior conversations, learn preferences such as a user's writing voice and edge cases, and gradually learn when they should interrupt for approval versus continue independently. SpaceXAI says they can later resume dropped threads, nudge stalled handoffs and pick up work from earlier conversations.

It further says Bots can become proactive over time, sometimes identifying work before the user explicitly asks for it. That is a more ambitious claim than conventional scheduled automation and will put additional pressure on permission controls and escalation rules if the system is deployed against production applications.

Bots can delegate work to other Bots

Grok Bot also supports multiple agents operating together.

Users can place several Bots into the same thread, where the agents can pass work between one another. The company's demonstration includes specialized Research, Communications, Chief of Staff and Travel Bots coordinating tasks.

SpaceXAI says those Bots can independently message one another and share context within threads. Users can also put multiple Bots into a group conversation where they assign ownership, transfer work and coordinate among themselves, bringing the human back in primarily for judgment calls.

Internally, the company says employees sometimes place a Chief of Staff Bot above specialist Bots responsible for functions such as inbox management, recruiting, expenses, operations and bug fixes. That makes the product's orchestration model more explicit: the user does not necessarily have to serve as the routing layer between every specialized agent.

Initial reactions are extremely positive

Lenny Rachitsky, host of the popular vlog and podcast Lenny's Podcast and author of newsletter Lenny Letter, received early access to Grok Bot and loved using it. As Rachitsy wrote on X : "I haven't been this excited about a new AI product in a while. It's like OpenClaw, but super easy, reliable, and less scary to use. I think this will be a huge new product line for Cursor/Grok/SpaceX."

Similarly Matt Shumer, an AI entrepreneur who said he tested Grok Bot for several weeks before launch, highlighted this orchestration as one of the product's strongest features.

“The best way I can describe it is an agent for everything, not just code,” Shumer wrote on X.

In one test, Shumer said he created separate researcher and writer Bots, then created a Chief of Staff Bot and instructed it to coordinate the other two on a project. He expected the workflow to break down.

“It worked out of the box,” he wrote.

His main criticism involved model selection.

Unlike systems where developers or advanced users explicitly select the underlying model, Shumer said Grok Bot automatically routes tasks to models on the backend.

“You don’t choose a model for your Grok Bot,” he wrote. “It’s all done automatically on the backend.”

Shumer said the model router “wasn’t great” during his testing, although he said he was subsequently told it had improved.

SpaceXAI's expanded announcement still does not identify which underlying models the router uses, nor does it document a mechanism for users to select, pin or switch to a particular xAI or third-party model. As a result, the model layer remains largely abstracted from users in the publicly supplied launch material.

That abstraction represents an important tradeoff for enterprise deployments. Automatic routing can remove a significant configuration decision for ordinary employees, but advanced users may want explicit control over model cost, latency, reliability and behavior — particularly for repeatable production workflows.

The agent market is moving toward longer-running work

Grok Bot enters a market increasingly focused on agents that can do more than generate text or code.

Anthropic's computer-use capability established a mechanism for Claude models to interact with software through screenshots, cursor movements, clicks and typing. Its broader Claude product also connects with workplace services and remote MCP servers.

OpenAI, meanwhile, now describes ChatGPT Work as an agent for “longer, multi-step work and finished deliverables,” while keeping Codex focused specifically on software development. OpenAI's enterprise agent economics can also incorporate usage-based credits, making task complexity and token consumption part of deployment cost calculations.

Grok Bot's differentiation is therefore less about proving that AI can operate software than packaging computer use, persistence, workflow learning and multi-agent coordination into something resembling a workforce interface.

SpaceXAI's announcement sharpens that distinction by emphasizing completion rather than assistance. One company product employee, identified only as Roman, describes the difference as closing the gap between work that is nearly finished and work actually completed inside the destination application: “Grok Bot can finish the swing, because the work lands where a human would put it, in the actual tool.”

That distinction will ultimately depend on reliability. A chatbot producing a bad answer creates a correction problem. An autonomous agent operating CRM records, support queues, vendor conversations or other production systems can create an operational problem.

Grok Bot's success will therefore depend not only on model intelligence, but also on permissions, predictable execution, escalation behavior, memory accuracy and how reliably agents recognize when human approval is necessary.

That challenge becomes more significant if Bots act proactively, resume forgotten work and coordinate with one another without the user serving as an intermediary. Those capabilities reduce the amount of supervision required when they work correctly, but they also expand the consequences of an incorrect assumption, stale context or improperly scoped permission.

The interface may matter as much as the models

Shumer described the product's interface as feeling like iMessage, an intentionally familiar metaphor for a system whose underlying architecture — autonomous computers, persistent memory, agent orchestration and automatic model routing — could otherwise be difficult for nontechnical users to configure.

SpaceXAI makes essentially the same usability argument in its launch announcement. Rather than asking users to construct workflows before getting started, it says users can simply message a Bot from a phone or desktop, hand it work and later continue the same conversation from either device.

That simplicity is part of the product strategy. Grok Bot is trying to hide much of the conventional machinery of automation — workflow builders, explicit integrations, agent routing and orchestration — behind an interaction model that resembles messaging a coworker.

That may prove to be the larger bet behind Grok Bot.

The AI industry has spent several years making models increasingly capable of using tools and completing multi-step tasks. Grok Bot attempts to turn those capabilities into an organizational abstraction people already understand: give someone a job, teach them how you work, and let them coordinate with the rest of the team.

If that abstraction proves reliable, the enterprise agent competition may increasingly shift away from which assistant produces the best individual response and toward which platform can most reliably manage fleets of agents performing ongoing work.

Presented by Rezolve Ai

When an AI assistant recommends a product or brand, it generates something valuable: a purchase-ready consumer with high intent and low friction in their decision. That consumer has already compared options, asked follow-up questions, and arrived at a conclusion. They want to buy.

What they encounter next is a commerce infrastructure that was not designed for them.

The gap between recommendation and purchase

The typical enterprise commerce stack was built for a specific model: a consumer who arrives at a brand's website through search or a direct link, navigates product pages, adds to cart, and completes checkout through a multi-step form flow. That model assumed the consumer would do the work of bridging their intent to the transaction. Most commerce systems still assume exactly that.

Agentic commerce breaks that assumption. When intent is generated outside the brand's owned environment, the handoff to transaction becomes a structural problem. Context doesn't transfer. Sessions don't persist. The consumer who asked an AI assistant for a recommendation and received one now faces the same friction-laden checkout process as someone who arrived with no prior intent at all.

Cart abandonment rates have remained stubbornly high for years. Baymard Institute research puts the average at 70%. That figure predates the agentic commerce era. As more purchase intent is generated through AI interfaces, and as the gap between that intent and a brand's transaction layer widens, the abandonment problem is likely to get structurally worse before it gets better.

What the current stack wasn't built to handle

The commerce infrastructure most enterprises operate today was assembled over two decades of incremental investment. Each layer added a capability: a search tool, a recommendation engine, a personalization layer, and a checkout system. Each was built to solve a specific problem within a human-initiated shopping journey.

None of it was built to receive intent from an AI agent.

When an AI system generates a purchase recommendation, it needs to do more than surface a product page. It needs to verify real-time inventory. It needs to apply pricing logic and promotional rules. It needs to respect brand policy around which products can be recommended together, which channels apply which discounts, and what the correct fulfillment path looks like for a given consumer. And it needs to do all of that without breaking the conversational context that made the recommendation possible in the first place.

Current commerce stacks can't do this reliably. The systems that hold the relevant data, inventory, pricing, order management, fulfillment, are not exposed in ways that AI agents can safely and accurately access. The result is a journey that starts with intelligence and ends with a broken experience: a link out to a product page, a generic checkout flow, and a consumer who arrived ready to buy and left without completing the transaction.

The conversion problem is an architecture problem

The industry has treated conversion optimization as a front-end problem for most of its history: better copy, cleaner checkout UX, fewer form fields, smarter retargeting. Those interventions were appropriate for the model they were built to serve.

The agentic commerce era introduces a different kind of conversion failure, one that front-end optimization cannot fix. When intent is generated externally, conversion depends on whether the back-end infrastructure can receive that intent, act on it accurately, and complete the transaction within the guardrails the brand has established. That is not a UX problem. It is an infrastructure problem.

Brands that are investing heavily in AI-powered discovery while leaving their execution layer unchanged are widening the gap between the promise AI makes on their behalf and the experience they can actually deliver. That gap has a cost, measured not just in lost transactions but in consumer trust that erodes each time the promise and the reality don't match.

Rezolve Ai commissioned research across 1,500 US consumers in January 2025 that found consumers who encounter friction immediately after an AI recommendation are significantly less likely to complete a purchase than those who encounter friction at the top of a traditional funnel. The implication is direct: AI raises the expectation bar at the moment of intent. Brands whose infrastructure cannot clear that bar are paying a conversion penalty they may not even know they're incurring.

What closing the gap requires

Closing the gap between AI-generated intent and completed transaction requires rethinking which layer of the commerce stack carries the most strategic weight in an agentic world. For most of the past decade, that weight sat with discovery and experience. The brands that invested most in search, personalization, and content won a disproportionate share.

In the agentic era, the weight shifts to execution. The brands that can reliably take AI-generated intent and turn it into a governed, accurate, brand-safe transaction will have a structural advantage over those whose infrastructure stalls at the handoff.

That is a different investment thesis than the industry has operated on. And most enterprise commerce roadmaps have not yet caught up to it.

Sponsored articles are content produced by a company that is either paying for the post or has a business relationship with VentureBeat, and they’re always clearly marked. For more information, contact sales@venturebeat.com.

Mistral AI wants to turn European AI sovereignty from a talking point into a product — one with a service-level agreement attached.

The French artificial intelligence company announced Tuesday a three-part expansion of its infrastructure business: regional inference endpoints that let customers choose whether their AI workloads run in Europe or the United States, a new "Priority Tier" backed by an uptime guarantee for mission-critical deployments, and a coalition of European enterprises making multi-year compute commitments that Mistral says will underwrite 200 megawatts of infrastructure across Europe by the end of 2027 — and a full gigawatt by the end of 2030.

In a move that may raise eyebrows among sovereignty purists, the company also said it will begin hosting third-party open models on its platform, starting with GLM-5.2 from Z.ai, the Chinese AI lab formerly known as Zhipu.

Taken together, the announcements mark a decisive shift in how Mistral positions itself. The company that built its reputation training open-weight language models is now selling something closer to critical infrastructure: assured capacity, regional control, and contractual reliability for enterprises and governments that want frontier AI without surrendering control over where it runs.

"When we spoke in June, the story was around how Mistral was building a full-stack AI offering," Timothée Lacroix, Mistral's co-founder and chief technology officer, told VentureBeat in an exclusive interview ahead of the announcement. "Today, the announcement is about strengthening one part of this infrastructure, which is the inference part."

That one part, it turns out, comes with a price tag measured in the tens of billions of dollars.

Inside Mistral's plan to build 1 gigawatt of European AI compute by 2030

The headline numbers deserve scrutiny, because they imply staggering capital requirements. Mistral currently operates less than 200 megawatts of capacity, according to the company. Details shared with VentureBeat show the near-term buildout resting on three sites: a 44-megawatt facility near Paris that became operational in the second quarter of this year, a 23-megawatt facility in Sweden built in partnership with EcoDataCenter using renewable energy and advanced cooling, and a 10-megawatt site in Les Ulis, France, that came online in the third quarter.

Getting from there to one gigawatt by 2030 is a different order of magnitude. Independent estimates suggest just how different: research firm Epoch AI calculates that a typical one-gigawatt AI data center requires roughly $38 billion in upfront capital expenditure, with servers and GPUs — not buildings or land — consuming the majority of the cost. Goldman Sachs Research pegs next-generation AI facilities at $15 million to $20 million per megawatt before accounting for the chips inside them.

Lacroix did not dispute the scale of the challenge. The investment required for a gigawatt of capacity "is a large investment that requires also a lot of scaling and revenue behind it," he said.

The urgency, in his telling, comes from a supply crunch that is about to get worse. "More and more, and especially around 2027 and 2028, we see that the demand for AI compute is exceeding what the market has to offer, especially in Europe," Lacroix said. McKinsey has estimated that meeting global AI demand could require $5.2 trillion in data-center capital expenditure by 2030 — and Europe, by most analyses, is starting from behind.

A company valued at a fraction of its American rivals cannot close that gap with venture capital alone. Which explains the most consequential — and most unusual — piece of Tuesday's announcement.

European Compute Units turn AI sovereignty into a five-year contract

Mistral is assembling what it calls an anchor group of enterprises whose long-term commitments will collectively finance infrastructure none of them could justify alone. Those commitments convert into "European Compute Units," or ECUs — a claim on Mistral-built capacity over multiple years that participants can spend on inference, training, model adaptation, or other AI workloads as their needs evolve.

If that structure sounds more like a power-purchase agreement than a cloud contract, that appears to be the point. Data-center financing increasingly resembles large infrastructure projects — gigawatts, substations, energy agreements — rather than traditional technology spending, and lenders want demand locked in before capital gets deployed. Mistral raised €830 million ($962 million) in debt earlier this year to fund its data center near Paris, TechCrunch reported in March, and pre-committed enterprise demand is exactly what makes that kind of financing repeatable at ten times the scale.

Lacroix was unusually direct about the mechanics. "The entire point of compute units is to have commitment," he said. "The goal is to have customers commit for around five years, or at least a long time." Asked what happens if a customer wants out early, he didn't soften the answer: "There is no getting out."

What makes a five-year, no-exit commitment palatable, he argued, is flexibility in how the capacity gets consumed. "Typically this can be spent on raw inference that you then feed through any other AI stack. It can be spent on raw compute as managed Kubernetes, and it can be spent at the very top with our full AI offering," he said. "My hope is that they will use it with our full-stack services and will love it."

The anchor group already includes some of Europe's industrial heavyweights. Amadeus CEO Luis Maroto said in a statement that "capacity, deployment control, and operating continuity become increasingly important for all enterprises." ASML chief Christophe Fouquet — whose company led Mistral's $13.4 billion (€11.7 billion) Series C last year — called building European AI capacity one of the few industrial endeavors that "will matter more to Europe's next generation," while Capgemini's Aiman Ezzat framed it as "a question of who shapes the future of European industry." CMA CGM chairman Rodolphe Saadé said the shipping group's Mistral deployment is "already under way among thousands of employees."

Commitments of that duration only make sense, of course, if the sovereignty being purchased is real. On that question, Mistral's announcement contains an asterisk worth reading closely.

The fine print on sovereign AI: what data can still leave Europe

The centerpiece product is Mistral Regional Endpoints, now generally available, which let customers pin inference and its associated processing to Europe or the U.S. Alongside it, the new Priority Tier — in public preview — offers committed service levels, custom rate limits, and an uptime SLA for mission-critical workloads.

Mistral claims it is the only European AI lab offering both a choice of processing region and an SLA-backed service tier, and Lacroix said a third option is coming: an endpoint "that stays on Mistral-controlled infrastructure, so on Mistral compute" — for customers who want their inference not just in Europe, but off hyperscaler hardware entirely.

Then comes the fine print. Mistral's own materials note that in-region inference remains subject to "limited, safeguarded transfers" to sub-processors that may sit outside the chosen region. Pressed on what actually leaves Europe, Lacroix pointed to the connective tissue of modern AI applications: tool calls.

"There are some tool services, like some tool calls, that might be hosted in places where we don't fully control this," he said, citing web search as an example. "A few of our web-search providers might not all be in Europe, and in that case, we need to potentially gate that capability."

His answer to the compliance question — would this satisfy a European bank or a defense ministry? — was that gating is the feature, not the bug. Capabilities that cannot be sourced in-region can be switched off entirely, restricted to certain users or workspaces, or, given sufficient demand, rebuilt with European providers. "Any capabilities that we don't find a provider for in Europe — if it needs to be done in Europe, we'll find some way to implement it or find ways to address it," Lacroix said.

For enterprise buyers, that is a more honest framing than most sovereignty marketing offers: full regional control is available, but the moment an AI agent reaches out to the open web, sovereignty becomes a configuration decision rather than a default. The same pragmatism runs through the announcement's most surprising line item.

Why Europe's open source AI champion is hosting China's GLM-5.2

A French national champion — one that has partnered with the French army and positioned itself as Europe's answer to American AI dependence — hosting a Chinese lab's model invites an obvious question. Lacroix's answer was disarmingly matter-of-fact.

"It's a great model. Everyone loves it. It's open weight, so there was no good reason for us not to do it, really," he said, noting that Mistral's own stack is already built on open-source software like Kubernetes.

On security vetting, he argued that open weights fundamentally change the risk calculus. "The risks in taking a new model, at the layer of the weights, are — at least in my opinion — rather limited," Lacroix said. "We checked basically all of the safety and compliance evals that we have. We'll control that model, its outputs, and what it does the same way we do any of our models. We have the same inputs and outputs and monitoring capabilities over all of it."

The strategic logic is worth unpacking. By hosting third-party open models under European regional controls and the same SLAs as its own, Mistral is repositioning itself from model vendor to sovereign distribution layer — the trusted intermediary through which any open model, regardless of origin, can be consumed by a regulated European enterprise that could never call a Chinese API directly. It is the "model garden" playbook the hyperscalers run with Bedrock and Vertex, executed on European soil with European guarantees.

Customers appear to be reading it that way. "Mistral allows us to run open models under strict regional controls and service commitments, making it easy for us to maintain data residency and compliance requirements," Matan Griberg, CEO of AI software-engineering company Factory, said in a statement.

Lacroix stressed the move is not a retreat from frontier training: the model Mistral had in training as of June "is still training, and we're still very excited about it," he said. But openness to rivals' models signals where the company now believes its moat lies — not in any single model, but in the infrastructure underneath all of them. Which makes its relationship with the world's most powerful infrastructure company all the more interesting.

How the multibillion-dollar Microsoft deal funds Mistral's independence

Hovering over every sovereignty claim is Mistral's deepening relationship with Microsoft. In July, the two companies announced a multibillion-dollar expansion of their partnership under which Microsoft will rent capacity from Mistral's European data centers to serve its own cloud and AI demand, while adding Mistral Medium 3.5 and OCR 4 to Microsoft Foundry, bringing Medium 3.5 to Copilot Studio, and enabling Mistral models on Azure Local for disconnected, customer-controlled environments. Mistral CEO Arthur Mensch told The Wall Street Journal at the time that two-thirds of Mistral's customers already work with Microsoft.

How does a company selling independence from U.S. hyperscalers square taking one on as its largest tenant? Lacroix described Microsoft not as a patron but as an anchor customer that de-risks the buildout.

"It allows us to scale different parts of the business differently by building infrastructure with Microsoft as a customer," he said. "We can scale that team, we can scale our infrastructure, and make sure that we can then, on the side of it, also build for ourselves and for our customers." He compared the arrangement to the neocloud playbook — companies that built businesses supplying capacity to the hyperscalers themselves. "As that part of our business resembles that of neoclouds, we're following the same thing."

It is a genuinely clever inversion: rather than renting American infrastructure, Mistral is renting infrastructure to one of America's largest companies, using Microsoft's demand to finance capacity that also serves European sovereignty customers. But the independence has limits no contract can engineer away — the GPUs filling Mistral's European data centers come overwhelmingly from Nvidia and other American chipmakers, as SiliconANGLE noted in its coverage of the July deal.

Asked directly why a customer should choose Mistral over an EU region on AWS or Azure, Lacroix gave two answers. "The simplest possible answer is capacity. There is more demand than supply right now, and so it adds another option," he said. The second cuts closer to the pitch: "We are a European provider, and on the region that would be Mistral compute, we are fully independent. That's a truly differentiated offering than all of the hyperscalers or pure inference companies can provide."

The economics of open models: why agentic AI is pushing inference to the cloud

There has always been a tension at the heart of Mistral's business: its best-known models are free to download, and open models have historically been difficult to monetize through APIs. Asked how free weights fund a gigawatt buildout, Lacroix offered the clearest articulation yet of the company's thesis — that the economics of self-hosting are collapsing under the weight of the models themselves.

"When the models were smaller, and we were before the explosion of agentic AI, it was doable for enterprises to host their own — up to, let's say, 100-billion-parameter dense models — on their premises," he said. "More and more, with models going into the trillion or more parameters, with the current hardware, and with the increasing amount of tokens that need to be processed, it becomes harder."

His conclusion was blunt: "I don't see how, with the current trend of model size and growth of agentic tokens, we keep the full inference on-prem. To me, that is why we think we're going to monetize our cloud inference." Inference, he noted, is particularly well suited to the cloud because it "does not need to hold any data" and can be encrypted in transit.

In other words: open weights get Mistral into the enterprise, and the physics of trillion-parameter agentic workloads brings the inference — and the revenue — back to Mistral's data centers. The thesis will get an expensive test. Mistral has raised roughly $4 billion to date, according to PitchBook data — a fraction of the war chests assembled by OpenAI and Anthropic — and Bloomberg reported in June that the company is in talks to raise about €3 billion at a roughly €20 billion valuation, nearly double its Series C mark. The revenue behind the buildout will have to come from exactly the enterprises Tuesday's announcement is courting.

And Europe, in Mistral's telling, is only the first market for what it is selling. Asked whether the framework could be replicated in the Middle East, Asia, or anywhere else anxious about AI dependence, Lacroix didn't hedge: "It's completely right. We're starting this in Europe because it's also an easier part of the world for us to scale into, especially in the infrastructure. But we definitely want to extend this, depending on customer demand." Every layer of the stack, he said, "can be controlled, changed, replaced depending on where we operate and what the requirements are — that's pretty much where we excel."

That is the wager underneath the SLAs, the compute units, and the Chinese model flying a European flag: in a world where the U.S. and China dominate frontier AI, the durable business is selling everyone else control. To fund it, Mistral is asking Europe's largest enterprises to sign five-year contracts with no exit — while making a bigger, longer commitment of its own. A gigawatt, after all, is a promise measured in decades. For Mistral, too, there is no getting out.