Enterprises deployed AI agents ahead of the controls needed to manage them — and they did it knowingly. That is the central finding across the five parallel surveys VentureBeat Research fielded in June, spanning every layer of the agentic stack. Now those enterprises are retrofitting to catch up with their own standards, and they are budgeting for it: In each of the five control layers we measured, 57 to 68% of enterprises plan to switch vendors or add new ones within 12 months, and roughly a third, depending on the layer, plan to move within the quarter.

VentureBeat Research measured the five controls an enterprise has to build before it can trust an agent: identity, evaluation, cost telemetry, the context layer, and orchestration. Identity governs which agent is allowed to do what, under whose credentials. Evaluation determines whether the agent's work is any good. Cost telemetry tracks what each agent costs to run. The context layer supplies the business data and definitions agents draw on when they answer. And the orchestration control plane coordinates multi-step agent work. Each of our five reports measures one of those controls.

Most deployed "agents" are chatbots wearing the label. Seventy-one percent of enterprises said a quarter or fewer of their deployed "agents" can complete multi-step work on their own; only 10% said true agents are the majority of what they run. These respondents are positioned to know: 81% recommend or decide AI purchases at their companies. A single-prompt chatbot with a human reading every answer needs none of the controls the other four reports measure. A true multi-step agent needs all of them — and most enterprises can't say which one they've deployed. (Full findings: Agentic Orchestration report.)

Autonomy is outrunning trust in the evaluations that gate it. Two-thirds of enterprises either already allow an agent to push a code or system change to production on automated evaluation results alone, with no human review, or are actively engineering toward that within 12 months. Only 5% fully trust the evaluations that would make that call — and half of enterprises shipped an agent that passed internal evaluations and then caused a customer-facing failure in the past year. Before removing human review from any workflow, test evaluations against production outcomes rather than internal benchmarks. (Full findings: Agent Reliability & Evals report.)

Companies that let agents share credentials get hit more often. Sixty-nine percent of companies let at least some of their agents share credentials — multiple agents operating under one API key or service account. Organizations that allow credential sharing anywhere experienced a security incident or near-miss at a 63.5% rate (47 of 74), against 40.9% (nine of 22) at companies where every agent has its own scoped identity. The fix is scoped identity for every agent, starting with the ones that touch production systems. (Full findings: Agentic Security & Identity report.)

The most expensive hardware in the building runs at half capacity or less. More than eight in 10 enterprises that run their own GPUs reported utilization of 50% or less, and only 44% rigorously track what their AI compute actually costs and returns. The number worth chasing first isn't more GPUs — it's the utilization and per-workload cost of the ones already running. (Full findings: AI Infrastructure & Compute report.)

Agents answer confidently from data nobody governs. Fifty-seven percent of enterprises traced a confident, wrong agent answer in the past six months to their own missing or inconsistent business context — wrong metrics, stale definitions, absent documents — and most saw it happen more than once. Governing the definitions agents answer from — metrics and entities first — has to come before scaling the agents that depend on them. (Full findings: Context Layers / RAG report.)

No layer has an entrenched incumbent: The defaults today are the built-in tools that ship with the big AI platforms enterprises already use. Switching intent runs highest in orchestration itself, where 68% plan to adopt, add, or replace platforms within 12 months and 34% within the quarter. Our surveys did not ask which direction that money moves — toward the platforms' built-in tools or toward the specialists challenging them — and that open question is the next four quarters of this market.

About this research

VentureBeat Research fielded five parallel surveys in June 2026 under its VB Pulse program: Agentic Orchestration (101 respondents), Agent Reliability & Evals (157), Agentic Security & Identity (107), AI Infrastructure & Compute (107), and Context Layers / RAG (101) — 573 qualified respondents in total, all at organizations with 100 or more employees. Samples are self-selected, and some findings should be read directionally; each report carries its full methodology note. What the pattern supports more strongly than any single percentage is the direction: every survey, independently, points the same way. VentureBeat produces both this research and VB Transform, the conference where these reports debuted.

Anthropic released Claude Opus 5 on Friday, a model the company says delivers nearly all the intelligence of its top-of-the-line Claude Fable 5 at half the cost — a launch that signals how the AI race is shifting from raw capability to the economics of daily use.

The model, available immediately on all of Anthropic's platforms, is priced at $5 per million input tokens and $25 per million output tokens, unchanged from its predecessor, Opus 4.8. It becomes the new default model on Claude Max, Anthropic's premium consumer tier, and the strongest model available on Claude Pro.

The positioning is deliberate. Anthropic is not claiming Opus 5 is its smartest model — that distinction still belongs to Fable 5, and rival systems retain an edge in certain domains. Instead, the company is making a subtler argument that may matter more to enterprise buyers: that the most economically important AI work happens in a middle band of difficulty, where near-frontier intelligence delivered efficiently and cheaply beats frontier intelligence delivered expensively.

"Opus 5 as your daily driver, the model you hand complex work to and review when it's done," an Anthropic spokesperson said in an interview with VentureBeat, describing how the company's lineup now stratifies. "Fable 5 for your most ambitious work, the days-long autonomous projects nothing could take on before... Sonnet 5 for work you run at scale, where speed and cost per call decide what ships. Haiku 4.5 for subagents and instant answers."

How Claude Opus 5 benchmark results stack up against Fable 5 and rival AI models

On paper, the results are striking. Anthropic says Opus 5 sets new state-of-the-art marks on coding and knowledge-work evaluations including Frontier-Bench and GDPval-AA. On Frontier-Bench v0.1, an agentic terminal coding benchmark, Opus 5 scores 43.3 percent — more than double Opus 4.8's 18.7 percent and well ahead of Fable 5's 33.7 percent — at a lower cost per task, according to the company. On ARC-AGI 3, an evaluation of novel problem-solving, Anthropic reports Opus 5 scored three times as high as the next best model. On OSWorld 2.0, a computer-use benchmark, the company says the model surpasses Fable 5's best result at just over a third of the cost.

The numbers come with honest caveats that are themselves notable in an industry prone to superlatives. Anthropic acknowledges Opus 5 remains behind Mythos 5, a competing model, on cybersecurity tasks and biology research, and an OpenAI-family model still leads on one agentic coding benchmark.

The more revealing caveat came from Anthropic itself, when asked where Opus 5 still falls short of Fable 5. The spokesperson's answer amounted to a candid admission about what benchmarks do and don't capture.

"The evals where Opus 5 wins are bounded tasks with a specific outcome, which is where it's strongest. What those evals don't measure is duration," the spokesperson told VentureBeat. "One way to put it: Opus 5 is the best tool for the jobs benchmarks can see, and Fable 5 is what you reach for when the job outruns the benchmark."

Fable 5, by contrast, "is for the longest, most autonomous jobs, where the model has to stay coherent across many connected steps over hours or days with dense source material," the spokesperson said, advising customers to "run both on a representative workload, one bounded task and one long-horizon job." That framing — bounded tasks versus long-horizon autonomy — may become the defining axis of model differentiation in 2026, as benchmarks saturate and the hardest remaining problems involve sustained, multi-day agentic work rather than discrete puzzles.

Why token efficiency is becoming the real battleground for enterprise AI spending

Threaded through the launch is a theme Anthropic clearly wants buyers to absorb: Opus 5 doesn't just score well, it scores well per dollar. The model ships with an adjustable "effort" setting that lets customers trade intelligence for speed and token savings, and Anthropic's charts emphasize performance at a given cost rather than peak performance alone.

Early customers echoed the point with unusual specificity. Harvey, the legal AI company, said Opus 5 achieved similar performance to Opus 4.8's maximum-reasoning mode "while generating 26% fewer tokens on average," according to Niko Grupen, its head of applied research. Richard Pham of Fundamental Research Lab said that on hard financial-modeling tasks, the model averaged nine percentage points higher accuracy "while using roughly one-third fewer turns and tool calls and 60% less time."

Wade Foster, chief executive of Zapier, said Opus 5 topped his company's AutomationBench leaderboard "without spending more tokens than prior Claude models," running a full churn-prevention workflow from start to finish. "Previous models didn't pass; Opus 5 hit 100%," he said. Scott Wu, chief executive of Cognition, the company behind the Devin coding agent, said that on FrontierCode 1.1, "Claude Opus 5 approaches Fable-level performance at half the cost," with particular strength in debugging and root-cause analysis.

The efficiency emphasis reflects commercial reality. Enterprise AI spending is no longer experimental, and inference costs — the price of actually running these models at scale — have become a board-level line item. 

Anthropic's business skews heavily toward API and enterprise usage; according to a February 2026 analysis by Contrary Research, Claude held roughly 40 percent of the enterprise large language model market by usage as of late 2025, and Claude Code alone had reached about $1 billion in annualized revenue. For a company whose customers pay by the token, a model that does more with fewer tokens is not a nice-to-have. It is the product.

Self-verifying AI agents and what they mean for the hidden costs of automation

Beyond the numbers, Anthropic is selling a behavioral story: that Opus 5 verifies its work and iterates until it succeeds. The company offered several examples from testing that read like small parables of machine stubbornness.

In one Frontier-Bench task, the model was asked to reconstruct a machine part as a 3D CAD model from a drawing it was intentionally given no way to view. Rather than fail, Anthropic says, Opus 5 wrote its own computer vision pipeline to extract the geometry from raw pixels — and did so repeatedly, while no competing model solved the task in five attempts. In another case, given a real bug in a popular open-source package manager, the model found the root cause and fixed an edge case the community's own patch had missed; a competing model patched only the symptom and declared victory. An engineer at a trading firm, the company says, used Opus 5 to build a market data feed for a new exchange in a single session and, finding no live feed to validate against, watched the model build its own test harness to check its parsing code.

Customers described similar behavior in the wild. Cristian Rivera, a staff software engineer at Stripe, said he gave the model "a chief-of-staff role over my dev environments" for a weekend: "it built its own monitor, drove each box, and pulled me in only for the judgment calls."

This is the capability enterprises actually care about, and it is worth dwelling on why. The gap between a model that produces plausible output and one that verifies its output is the gap between a demo and a deployable system. Most of the hidden cost of enterprise AI today is human review — engineers checking the machine's work. A model that reliably checks its own work compresses that cost, which is precisely why customers keep citing fewer turns, fewer passes, and less time rather than higher raw scores.

Inside Anthropic's safety strategy: capability gaps, classifiers, and model fallbacks

The launch also showcases Anthropic's increasingly intricate approach to safety — one that now involves deliberately not teaching its models certain skills. The company says its automated behavioral audit found Opus 5 to be its most aligned model to date, scoring 2.3 on overall misaligned behavior, lower than Opus 4.8, Sonnet 5, or Fable 5, with the lowest rates of deceptive behavior and the least susceptibility to being tricked into misuse.

On the capability side, Anthropic says it intentionally avoided training Opus 5 on cyber tasks, as it did with Opus 4.8. The model improved on them anyway — a side effect of general capability gains — and now nearly matches Mythos 5 at finding software vulnerabilities. But it remains far behind at exploiting them: on Anthropic's OSS-Fuzz evaluation, Opus 5 identified vulnerabilities at a 79.4 percent rate, close to Mythos 5's 80 percent, but succeeded at developing exploits in only 4 challenges versus Mythos 5's 13. That asymmetry — strong at defense-relevant discovery, weak at offense-relevant exploitation — appears to be by design, and the safeguards follow the same logic. Anthropic expects Opus 5's cyber classifiers to intervene about 85 percent less often than Fable 5's.

When a classifier does trigger, requests in Claude.ai, Claude Code, and Claude Cowork fall back to Opus 4.8 by default — raising an obvious question: if a request is too risky for one model, why is it acceptable for another? "The model it falls back to has lower capability levels making the risk of harmful use lower as well," the spokesperson said, adding that "there is a message that lets the user know when this occurs and is visible in the chat."

The logic is defensible, but it reveals how AI safety actually works in 2026: risk is not a property of the question alone, but of the question multiplied by the capability of the system answering it. On biology, the calculus runs the other way. Opus 5 is now Anthropic's most capable generally available model for scientific research — scoring 10.2 percentage points higher than Opus 4.8 on the company's internal chemistry benchmark — though the spokesperson acknowledged that "Mythos 5 remains the stronger model for long-horizon, open-ended work like autonomous drug design campaigns."

The business stakes behind the launch: a $380 billion valuation and massive compute bets

The launch lands at a moment of extraordinary commercial momentum — and extraordinary obligations — for Anthropic. Reuters reported in February that the company was valued at roughly $380 billion in its latest funding round, following a period in which, per Contrary Research's analysis, its annualized revenue climbed from about $1 billion at the end of 2024 to a projected $9 billion by the end of 2025, with internal targets reportedly reaching $20 to $26 billion for 2026. Those targets are underwritten by enormous infrastructure commitments, including a reported $30 billion Azure compute deal alongside arrangements with Google Cloud and Nvidia — spending that only pencils out if enterprises keep expanding usage.

That is the context in which Opus 5's pricing strategy makes sense. Holding the price at Opus 4.8 levels while roughly doubling performance on key agentic benchmarks is effectively a steep price cut per unit of capability, designed to widen the funnel of workloads that are economical to automate. Every task that was marginal at Opus 4.8's cost-per-success becomes viable at Opus 5's — and every viable task is recurring token revenue.

The regulatory backdrop has grown more complex as well. A U.S. judge gave final approval this week to Anthropic's $1.5 billion copyright settlement with book authors, Reuters reported, closing a chapter of litigation over the company's early training data. And in June, Reuters, citing Axios, reported that the U.S. government had moved to block foreign access to Anthropic's most advanced models — a reminder that frontier AI is now entangled with export policy in ways that shape which customers can buy what.

Also shipping Friday: a Fast mode running at roughly 2.5 times default speed at twice the base price, automatic fallback routing on the API, and mid-conversation tool changes that no longer invalidate the prompt cache — a small feature that agent developers may appreciate more than any benchmark. Consistent with prior Opus models, Opus 5 carries no data retention requirements for general access, a point the spokesperson flagged unprompted for customers with "a hard zero data retention requirement." Developers can access the model as claude-opus-5 on the Claude API starting today.

Two questions will determine whether the bet pays off: whether Opus 5's efficiency claims survive contact with production workloads at scale, and whether enterprises embrace a world where safety classifiers, not users, sometimes decide which model answers. But the deeper message of Friday's launch is that the AI industry's center of gravity has moved. For three years, the labs competed on what their best model could do on its best day. With Opus 5, Anthropic is competing on something less glamorous and far more lucrative: what a very good model can do every day, for half the price. In a market where the frontier keeps moving, Anthropic is wagering that the real fortune lies just behind it.

Microsoft AI released two new in-house models into public preview on Wednesday — MAI-Image-2.5-Pro, its highest-fidelity image generator to date, and MAI-Voice-2-Flash, a speech model built for high-volume enterprise workloads — while publishing production data that amounts to the company's most aggressive argument yet that it can power its own products without leaning on OpenAI's frontier models.

The announcement, made by Microsoft AI's Superintelligence team, lands roughly a year after the company committed to building purpose-built models internally, and it arrives with an unusual level of specificity about where those models now run: Bing, PowerPoint, OneDrive, Dynamics 365, Excel, GitHub Copilot, and Azure. The message to enterprise buyers — and, implicitly, to OpenAI — is that Microsoft's homegrown models are no longer research projects. They are production infrastructure serving millions of users.

"Each of these enhancements is a step toward the same goal: Microsoft products, powered by Microsoft models," the company wrote in its announcement blog.

How MAI-Image-2.5-Pro and MAI-Voice-2-Flash stake out opposite ends of the AI cost curve

The two new releases occupy opposite ends of what Microsoft calls the quality-speed-cost curve, and the positioning is deliberate. MAI-Image-2.5-Pro targets the premium tier: hero imagery, detailed editing, and precise in-image text rendering — the last of which has long been a notorious weak spot for image generation models. Microsoft priced the model at $5 per million text input tokens, $8 per million image input tokens, and $106 per million image output tokens. The base MAI-Image-2.5 model recently launched at No. 2 for image editing on Arena, the community leaderboard that has become a de facto scoreboard for generative media.

The creative industry appears to be taking notice. Rob Reilly, global chief creative officer at advertising giant WPP, called the Pro model "a strong leap forward for GenMedia tools" in a statement included in Microsoft's announcement, adding that "Microsoft has firmly established itself among the leaders in generative AI."

MAI-Voice-2-Flash goes the other direction. First previewed at Microsoft's Build conference, Flash runs twice as fast as MAI-Voice-2 and costs 32% less, priced at $15 per million characters. It is designed for the unglamorous but enormous market of high-volume voice — call centers, voice agents, and real-time speech applications where latency and cost-per-call matter more than marginal gains in expressiveness. Together, the two models reflect a strategy of building families of models rather than a single flagship, because, as the company put it, a creative studio chasing maximum fidelity has very different needs from a customer service operation handling millions of calls a day.

Microsoft's production metrics show in-house models cutting GPU costs by up to 89%

The model launches are arguably less newsworthy than the deployment metrics Microsoft attached to them — numbers that read like a systematic case for swapping out third-party frontier models across its product portfolio. 

Bing Image Creator now runs entirely on MAI-Image-2.5, end to end, marking the first time the consumer image tool is fully in-house. In PowerPoint, Microsoft says MAI-Image-2.5 reduces GPU costs by up to 84% compared with GPT-Image-2, OpenAI's image model. In OneDrive, where MAI-Image-2.5 is now the default for key image-editing scenarios, the company reports a 26% increase in save rates, roughly 25% lower P95 latency, and 2.5 times greater efficiency under medium-utilization production workloads.

On the voice side, MAI-Voice-2-Flash now powers Dynamics 365 Contact Center — the platform used by customers including T-Mobile and EasyJet — where Microsoft claims GPU cost reductions of up to 89%. The model is also integrated into Azure Voice Live for developers building speech-to-speech agents.

Perhaps the most consequential deployment sits in healthcare. Microsoft's Dragon Copilot, used by 170,000 medical providers and responsible for processing 28 million patient encounters last quarter, now runs on MAI-Transcribe-1.5 for its multilingual workflow across 58 languages. Microsoft says internal evaluations show a 50% relative reduction in both transcription and language-identification error rates across most languages — a meaningful claim in a domain where transcription errors can propagate directly into clinical notes.

Inside the 'hill-climbing' strategy that lets small models beat GPT-5.6 in Excel

In a companion post published the same day, Microsoft detailed the methodology behind these results — what it calls its "hill-climbing machine," an integrated flywheel of data, models, and the product "harness" that surrounds them.

The clearest example is MAI-Code-1-Flash, the lightweight coding model launched in GitHub Copilot in June. Microsoft says the model achieves an approximately 10% higher code accept rate than GPT-5.4 Mini and Claude Haiku 4.5 in VS Code, while using 10% fewer median tokens. Developer retention tells a similar story: users were 6% more likely to return across multiple days than with GPT-5.4 Mini, and 11% more likely than with Claude Haiku 4.5.

Then Microsoft did something more interesting. It took the MAI-Code-1-Flash checkpoint and further trained it inside an Excel reinforcement learning environment, teaching a coding model the tools and workflows of spreadsheet knowledge work. The result, according to production user feedback, is a model on par with GPT-5.6 for the most common Excel tasks — while being small enough to run on Nvidia's older H100 and even A100 GPUs rather than requiring the latest-generation accelerators.

That hardware detail deserves emphasis. Every major AI company is fighting for allocation of cutting-edge chips, and a model that delivers frontier-adjacent quality on two-generation-old silicon fundamentally changes the deployment economics. It also frees the newest hardware — including Microsoft's now-operational GB200 cluster — for training rather than serving.

Satya Nadella's 'frontier diffusion' manifesto redraws the OpenAI relationship

Microsoft CEO Satya Nadella framed the announcements in a lengthy post on X titled "Frontier Diffusion & Control," which functions as something close to a strategic manifesto. "We can now take saturated frontier capabilities and deliver them at scale and at lower cost through models optimized for high-usage products, while continuing to use frontier models for frontier needs," Nadella wrote, adding that Microsoft is "beginning to route traffic across our first-party surfaces to MAI whenever our models match or outperform frontier alternatives."

Translated from executive prose: capabilities that were state-of-the-art a year ago are now table stakes, and Microsoft believes it can replicate them cheaply for the specific, repetitive tasks that dominate real product usage. Why pay frontier prices for a frontier model when a user just wants to reformat a spreadsheet column?

Nadella was careful to note that "frontier models from OpenAI and Anthropic are part of the orchestration system alongside MAI" — but he also articulated a pointed principle of model independence, arguing that a company's evaluations "should continue to hill climb even when any given model has been removed." 

“Keeping the harness, memory, context, and skills outside the model, he argued, is what gives Microsoft control. The subtext is hard to miss. Reuters reported in April that Microsoft’s exclusive license to OpenAI’s technology had been revised into a non-exclusive arrangement, and The Information reported last September that Microsoft had begun incorporating Anthropic models into some products. Wednesday’s announcement completes the triangle: Microsoft as orchestrator, with its partners’ frontier models as interchangeable components and its own models absorbing an ever-larger share of routine traffic.”

Developers cheer cheaper task-specific models while skeptics question Microsoft's track record

The response online captured both the appeal and the skepticism surrounding the strategy. "I love when people use small models for niche tasks," wrote one X user, @mavihsk, responding to Nadella's post. "Why do I have to use the all-knowing model just to change my field in Excel?" Another user, @nabu_lines, distilled the pitch neatly: "cost and performance both improve when you stop overusing the biggest model."

Others were less charitable about Microsoft's execution track record. "Microsoft is the worst when it comes to listening to user feedback," wrote designer @designedbyabin, arguing the company "will lose the AI race because they repeatedly failed to understand user needs." And one user, @tokenoverflow, offered a drier critique of the model-independence pitch: "i want it keep hill climbing after removing microsoft."

The skeptics raise a fair point. Microsoft's self-reported metrics — accept rates, save rates, GPU savings — come from its own internal evaluations, not independent benchmarks, and the company chooses which comparisons to publish.

But the strategy's logic does not depend on any single number. Nadella's framing that software now has "real marginal cost for the first time" explains why Microsoft is obsessive about tokens, GPUs, and serving costs: when AI features run on every keystroke across a billion-user product portfolio, an 84% GPU cost reduction is not an optimization. It is the difference between a viable business and a money pit.

Why Microsoft is turning its internal AI playbook into an Azure product

The final piece of the strategy is that Microsoft is selling the playbook, not just the models. Nadella explicitly positioned the hill-climbing approach as "a template for every other AI native, SaaS, or Enterprise company," and Microsoft is packaging the toolchain through Foundry and what it calls Frontier Tuning — letting enterprises train specialized models against their own proprietary evaluations and reinforcement learning environments. That turns Microsoft's internal cost-cutting exercise into an Azure product, and it gives enterprise customers a reason to run their AI workloads on Microsoft's cloud even if the models themselves come from elsewhere.

The company's emphasis on models trained "on clean, traceable, enterprise-grade data, without distillation from third-party models" serves the same commercial end. In an industry facing mounting scrutiny over training data provenance, Microsoft is betting that enterprise buyers — and courts — will care where model capabilities come from. Microsoft says it is now extending the hill-climbing approach to Copilot Chat, Outlook, and PowerPoint, and both new models are available in public preview through Microsoft Foundry and the MAI Playground. "None of this is an endpoint," the company wrote. "We're just getting started."

Seven years ago, Microsoft bet more than $13 billion that OpenAI would build the future of AI. Wednesday's announcement suggests the company has since learned a cheaper lesson: the future of AI may belong to whoever builds the frontier, but the profits belong to whoever makes it ordinary.

Two weeks after debuting its more naturalistic GPT-Live audio AI model with full-duplex capabilities (listening and speaking at the same time), OpenAI is bringing it directly into developer workflows.

The company announced that GPT-Live now powers the ChatGPT desktop application on macOS and Windows, integrating directly with agentic systems like Codex and ChatGPT Work (which are separate experiences available in the ChatGPT desktop app).

When OpenAI initially launched GPT-Live on July 8, 2026, it introduced a continuous audio model capable of listening and speaking simultaneously—eliminating rigid turn-taking while delegating complex reasoning to background models like GPT-5.5.

Today's release expands that conversational layer to technical tasks, enabling software engineers to orchestrate multi-threaded coding jobs, review pull requests, and debug applications using natural voice commands.

As such, it could usher in a new era of "hands free" software development and even live, in-person group coding parties for the more than 10 million weekly active users across Codex and ChatGPT Work. Codex, of course, is the name given to OpenAI's models and harness focused on coding, but which the company has this year expanded into a more general productivity platform. An OpenAI spokesperson told VentureBeat this is the first time voice activation

OpenAI posted a promotional video showing some of its employees, Codex developer experience engineer Jason Liu and Codex technical staffer Guinness Chen, speaking to the same ChatGPT desktop app session in the same room, each issuing different instructions and conversing with the same model.

New capabilities unlocked

At its core, this integration relies on decoupling the real-time voice layer from the underlying execution engines.

While GPT-Live maintains fluid conversation—inserting natural verbal acknowledgments like "got it" without interrupting the user—it passes heavy computational workloads to background reasoning models.

On macOS, the desktop application incorporates "Appshots" and screen context features, allowing ChatGPT Voice to analyze the frontmost window alongside local files, codebase structures, and active plugins.

This architecture creates a pair-programming dynamic where developers talk through problems conversationally while agents execute tasks asynchronously.

Rather than manually stopping coding sessions to type detailed instructions or switch windows, developers direct the system hands-free.

The full-duplex engine dynamically decides when to speak, pause, or invoke tools, maintaining conversational state even as background agents process complex code modifications.

Directing coding and complex builds with your voice alone

The central operational capability in this update centers on multi-task execution across Codex and ChatGPT Work environments.

Software engineers can initiate multiple concurrent task threads from a single spoken prompt. For instance, a developer preparing to ship a feature can instruct the system to investigate an open authentication bug, review a pending API migration pull request, and generate missing unit tests simultaneously.

The desktop application coordinates these actions across disparate contexts, tracing issues through Slack conversations, GitHub repositories, and local codebases.

Developers can also verbally convert design mockups into working code, splitting tasks across frontend, backend, and testing layers.

With support for multi-folder projects (build 26.715) and remote execution via iOS, engineers can check task progress, answer agent prompts, and redirect active jobs without switching applications or managing individual processes line by line.

Proprietary license

OpenAI’s voice-enabled desktop release operates under a proprietary, commercial enterprise model. Access is restricted to paid subscribers across Plus, Pro, Business, Enterprise, and Education plans.

For individual developers and corporate engineering departments, this commercial structure means the model weights, voice processing pipelines, and agent state architectures remain fully closed.

Organizations cannot modify or self-host the underlying systems. Furthermore, tasks initiated via ChatGPT Voice consume standard usage allocations directly from existing Codex and ChatGPT Work plan quotas, treating voice-triggered actions identically to standard agentic workloads.

Community reactions

Developer communities immediately noted the implications of bringing continuous full-duplex voice to autonomous coding workflows.

Reacting to the build 26.715 release announcement—which details voice integration and multi-folder project support—AI Insider journalist @ChrisGPT noted on X: "Today OpenAI will release voice and remote guidance for codex ! One step closer to personal AGI".

Early technical feedback highlights widespread enthusiasm for orchestrating complex agentic tasks hands-free, particularly when stepping away from the workstation or managing build pipelines remotely.

Black Forest Labs (BFL) is expanding its FLUX family beyond image generation with today's launch of FLUX 3, a multimodal frontier model trained to understand and generate images, or combined audio/video clips up to 20 seconds from a single prompt — and to extend the same underlying architecture to robotic vision and actions.

The Freiburg, Germany-based AI lab says FLUX 3 is jointly trained across those modalities rather than assembling separate image, video and audio models behind a common interface.

That distinction is central to the company's pitch: BFL wants enterprises to think about creative generation, simulation, computer use and robotics as connected applications of a single capability it calls visual intelligence — models, in the company's words, "that can perceive, predict, and act across physical and digital environments." This release marks BFL's first public video generation model.

FLUX 3 will be offered through four product lines: FLUX 3 Video, FLUX 3 Image, FLUX 3 Action and the upcoming, open source FLUX 3 Dev. FLUX 3 Video, with optional native audio generation, and FLUX 3 Action are entering a gated "Early Access" program now, to which anyone can apply, but which BFL must approve.

There is presently no public access through BFL's application programming interface (API) or those of partners yet, but the company says FLUX 3 Image will roll out in the coming weeks, followed by general availability. The limited initial availability rollout echoes the release strategies of new models from other frontier labs in the U.S. lately, including Anthropic and OpenAI, though those were ostensibly for security concerns and due to government request.

What the company has not announced is pricing, production service-level commitments, evaluation methodology, sample sizes, rater counts or any image-model benchmarks at all. Enterprise buyers therefore cannot yet calculate total cost of ownership or independently reproduce the video comparisons.

Another big notable omission: FLUX 3 is not launching with downloadable weights at this time, nor an open source license. BFL says faster and open-weight versions will arrive later this year, and its technical blog names FLUX 3 Dev as "open-weight access to a multimodal backbone, for content creation (video, audio and image) and action prediction" — a considerably broader commitment than any previous FLUX Dev release, all of which covered images only.

But it arrives last in the sequence. Developers accustomed to receiving a locally deployable FLUX variant alongside — or soon after — a major model announcement will have to wait. That delay does not negate the company's commitment, but it is disappointing given the role open weights have played in FLUX's adoption thus far.

Flux 3 is rated higher than the competition, but missing pricing and benchmarking details may prevent rapid enterprise adoption

BFL has published several benchmark comparisons, but they're qualified as preliminary — with full benchmark results and methodology to be published later during broader general availability.

In early head-to-head preference testing on 10-second, 720p text-to-video clips with audio, the company says FLUX 3 was preferred over Luma Ray 3.2 in 93% of comparisons, Runway Gen-4.5 in 77%, Grok Imagine Video in 69%, Kling v3 Pro in 60%, Happy Horse v1 in 59%, Happy Horse 1.1 in 57%, and both Seedance 2.0 and Google's Gemini Omni Flash in 52%.

One caveat travels with every one of those figures, and it comes from BFL itself. The chart carrying the results is labeled a "preliminary evaluation of an early FLUX 3 candidate" — meaning the numbers describe a pre-release checkpoint rather than the model now entering early access. That cuts both ways: the shipping model may perform better, but nothing published today measures what customers will actually call.

Luma Ray 3.2 and Runway Gen-4.5, where FLUX 3 posted 93% and 77%, are the softest comparisons on the list — established products, but not the models currently setting the pace in independent video rankings. Those are real wins, and they are the ones least likely to change an enterprise shortlist.

Seedance 2.0, at 52%, is a statistical coin flip against a model most Western enterprises cannot currently procure. ByteDance indefinitely postponed Seedance 2.0's international rollout after Netflix, Warner Bros., Disney, Paramount and Sony sent legal threats over alleged systematic copyright infringement, and that suspension remains in place. Tying a frozen product is neither a strong claim nor a damaging one.

Gemini Omni Flash, also at 52%, matters much more. Omni is the closest large-platform analogue to what FLUX 3 is attempting — multimodal input, video and audio-aware creation, conversational editing — and by BFL's own measurement, the two are indistinguishable on 10-second text-to-video quality.

Google's advantage in that matchup is that Omni is generally available via Google's Gemini API for $0.10 per second of generated 720p video, or a 10-second clip for around.

One regional wrinkle matters for a German company's home market. Editing uploaded video is unavailable to Omni Flash users in the European Economic Area, Switzerland and the United Kingdom, though editing video the model itself generated is permitted. A European enterprise that wants to run its existing footage through a generative editing pass cannot currently do so on Omni Flash.

Here's a rough guide for enterprises considering which video models to rely upon:

One architecture for media generation and physical action

FLUX 3 builds on Self-Flow, BFL's method for aligning multimodal understanding and generation within one architecture, publicized back in March 2026.

The company says it significantly scaled up compute and data to train across video, images and audio simultaneously, and that testing showed video generation and action prediction do not require separate foundations — the same architecture could be extended to action prediction without sacrificing what it learned from video.

"We place vision at the center of our approach because it is the most signal-rich medium of the physical world. Images convey structure, images and video teach spatial relationships, video teaches dynamics, and actions reveal causal relationships. But vision alone is not the complete picture," said Robin Rombach, co-founder and CEO of BFL, in a pre-release statement provided to VentureBeat. "True intelligence means perceiving the world: predicting how it will change, taking action, and learning from the results. Joint training within one unified architecture is what will get us there, because each training modality strengthens the others. Audio conveys timing, prosody, and physical events that elude vision. Language conveys goals, abstractions, and instructions that pixels cannot easily express."

He put the case more bluntly elsewhere in the announcement: "You can't cheat reality. A model that only learns images can only generate images. But the world is not made of still frames. It moves, sounds, changes, and responds."

BFL says FLUX 3 targets creative tooling, media, design, e-commerce and physical AI, supporting video generation with synchronized audio, precise image editing, product and material consistency across motion, multilingual generation and robotic action prediction. It is already being tested by Canva, Burda, Magnific (formerly Freepik), Krea and Picsart.

For creative software companies, the appeal is consolidation. A single foundation could potentially support storyboarding, image editing, product rendering, video variation and localization without repeatedly translating assets and instructions between disconnected models.

For robotics teams, the potential value is data efficiency. Models that already encode motion, object behavior and physical change may need less task-specific robot training than systems starting from raw demonstrations.

What FLUX 3 Video can actually do

The video tier is the most concretely specified part of the launch, and it settles a question that had been circulating as rumor: FLUX 3 generates clips of up to 20 seconds with audio in a single generation.

Every video output comes with native audio. For comparison, HappyHorse 1.0 tops out at 15 seconds of 1080p with synchronized audio — though BFL has not stated what resolution its 20-second clips run at, and its published evaluations were conducted at 720p. Still, a 20-second long clip from a single prompt is among the longest yet achieved, matching OpenAI's discontinued Sora model.

The capability list BFL published covers:

• Text-to-video generation.

• Image-to-video generation, either animating from a starting frame or using images as visual references.

• Video-to-video generation from a reference clip, carrying elements such as a specific character into a new scene or context.

• Generative video-audio continuation from existing video and audio input.

• Keyframe-to-video generation for controlled transitions between defined moments.

• Multilingual dialogue.

• A broad range of visual styles and aspect ratios, from candid camcorder footage to animation and cinematics.

• Typography generation and animated design.

• Agentic chaining of individual clips into longer, multi-shot sequences.

That last item is the one enterprise video teams should look at hardest. BFL claims the capabilities combine to produce sequences lasting several minutes, with visual references keeping characters consistent across scenes. If that holds up under production conditions, it addresses the constraint that has kept generative video out of most commercial pipelines: not clip quality, but continuity across shots.

It is also the capability where competition is most direct. HappyHorse 1.1's headline upgrade is R2V, or Reference-to-Video, which accepts multiple character reference images to hold identity stable across generated footage — the same problem, approached at the input layer rather than through agentic clip chaining. Alibaba also claims zero-drift lip sync and has specifically targeted the artifacts that mark commercial AI video as synthetic, including facial oiliness and over-sharpening. Character consistency is where this category is being contested, and both companies know it.

BFL says FLUX 3 Video is already particularly strong at human facial expressions, associating sounds with physical events, and multilingual output. On the image side, the company says preliminary evaluations conducted during midtraining show significant improvement over earlier FLUX versions in complex prompt handling and text generation, including high-accuracy text in multiple languages. It published no image benchmarks or win rates.

FLUX-mimic tests whether video models can become robot models

BFL is applying its unified-architecture thesis through FLUX-mimic, a video-action model built on FLUX 3 and developed with Swiss firm Mimic Robotics, one of the first partners to receive early access.

The technical blog describes two distinct routes to action prediction: integrating native action prediction directly into FLUX 3, scaling up the initial Self-Flow work; and using the pretrained video backbone as a dynamics-aware foundation from which specialized action models can be finetuned with limited task-specific data. FLUX-mimic is the second route — the FLUX 3 backbone combined with mimic's robot-learning and production-deployment expertise in dexterous manipulation.

FLUX-mimic is designed for general-purpose robotic manipulation: helping robots understand a visual scene, predict the consequences of an action, and adapt to new tasks with far less task-specific data.

BFL and Mimic Robotics say that depending on task difficulty, the model can be finetuned for a specific manipulation task with as little as 30 minutes of robot data, where prior approaches have required 30 or more hours.

"The hardest part of robotics is data," said Elvis Nava, CTO of Mimic Robotics, in a statement provided to VentureBeat. "Every new task normally means hours of a robot repeating itself. Because FLUX-mimic is built on top of frontier video models that already understand how the physical world behaves, it picks up a new task in minutes, not days. This way, we can leapfrog the current state of the art in robot learning."

BFL argues that a model trained only on images cannot understand a world that "moves, sounds, changes, and responds," and that physical understanding is what produces convincing generated footage. Google makes a nearly identical claim for Gemini Omni.

Its developer documentation cites "world knowledge" that combines "an understanding of physics" with Gemini's grasp of history, science and cultural context. Its marketing is blunter still: "Most AI models just predict the next pixel to build a narrative or an image. Gemini Omni is different," the company posted in June, crediting the model with "an intuitive understanding of forces like gravity, kinetic energy, and fluid dynamics for more realistic movements that follow real-world logic."

The practical consequence for enterprise buyers is that world-model language is not a differentiator. Two of the three leading video systems now market physical understanding as their central advantage, and neither has published a benchmark that measures it.

There is no standard test for whether generated water behaves like water, whether a dropped object falls at a plausible rate, or whether a sound arrives when the impact does. Human preference ratings capture some of it indirectly. Nothing else on offer captures it at all.

Open weights helped make FLUX an industry standard

BFL officially launched in summer 2024 and gained a name for itself in the AI industry in the intervening two years for its commitment to open sourcing high-quality AI image models beloved by developers, creatives, and enterprises.

The company's founders, including Rombach, Andreas Blattmann and Patrick Esser, previously helped create VQGAN, latent diffusion and Stable Diffusion, the latter the open source technology that kicked off broad AI generation capabilities for the masses and currently used by many AI image generators and companies.

That reach translated into commercial distribution. FLUX models now power generative features inside Adobe Photoshop, Picsart and Nous Research's Hermes Agent, among other platforms, and the company cites film director Martin Scorsese among professional users.

Wired magazine described Black Forest Labs as a relatively small company that nevertheless became a leading competitor to Silicon Valley's largest AI labs, with FLUX models ranking near the top of image benchmarks and becoming some of the most downloaded text-to-image models on AI code sharing community Hugging Face. The company says it now runs a 100-person team across Freiburg and San Francisco.

FLUX.1 Dev, FLUX.1 Kontext Dev, FLUX.1 Fill Dev and related control models, released shortly after the firm's launch, gave researchers and creative-tool developers access to downloadable checkpoints, local inference and integrations with frameworks including Hugging Face Diffusers and ComfyUI. FLUX.1 Kontext Dev, for example, was released as an open-weight model for research and noncommercial use, with generated outputs permitted for commercial purposes under the applicable license.

The company continued that pattern with FLUX.2 Dev in late 2025, a 32-billion-parameter open-weight model combining generation and multi-reference editing. Black Forest Labs called it the strongest open-weight image generation and editing model available at launch and released weights, reference inference code and optimized implementations for consumer Nvidia GPUs.

FLUX 3 Dev raises the stakes on that evaluation. Previous Dev releases were image models. This one is described as a multimodal backbone spanning video, audio, image and action prediction — meaning a single license will govern whether a company can locally deploy a model that touches both content production and physical machinery. BFL hasn't yet shared information about its license, the parameter count, quantizations or hardware requirements.

The company frames open weights as an enterprise feature rather than a community gesture, arguing they enable secure, low-latency local deployment for applications like robotic control systems and let teams adapt FLUX 3 to their own data, products and workflows.

The financial backing behind FLUX 3 is worth noting alongside the technical claims. Black Forest Labs is valued at $3.25 billion and has raised more than $450 million from investors including a16z, AMP, Salesforce Ventures, Nvidia, General Catalyst, Adobe Ventures, Figma Ventures, Canva and Deutsche Telekom's T.Capital.

When Cisco ran 6,986 multi-turn attacks against 15 flagship models, attackers who adapted across the conversation broke through as often as 88.3% of the time. Amy Chang, Cisco's head of AI threat intelligence and security research, brought that finding to the agentic security panel at VB Transform 2026; the number should worry anyone still running single-turn red-teaming programs.

VentureBeat's June 2026 Pulse survey of 107 enterprise respondents explains why the room was full. More than half, 54%, have already had a confirmed agent security incident (18%) or a near-miss caught before harm (36%). Just 32% give every agent its own scoped, managed identity, and fewer still, 30%, isolate their highest-risk agents in sandboxes. Provider-native and hyperscaler controls remain the primary agent security layer at 82% of companies surveyed. The world's largest security vendors have done the same math.

Palo Alto Networks closed its $25 billion acquisition of CyberArk in February, CrowdStrike agreed in January to pay $740 million for SGNL, and Cisco announced its intent to acquire Astrix Security for a reported $400 million, all of it aimed at the identity and isolation layer most enterprises have not finished building.

Chang came to the panel with almost two decades of experience spanning cybersecurity operations, government, and the military. She ran global cybersecurity operations as an executive director at JPMorgan Chase, where she led the bank's cyber threat intelligence teams, and served as a senior staffer on the House Foreign Affairs Committee and as a U.S. Navy Reserve officer. She also teaches cybersecurity and emerging threats as adjunct faculty at the Middlebury Institute of International Studies.

Chang's 88.3% number comes from a study she co-authored with Nicholas Conley, built on 30,090 single-turn prompts and 6,986 multi-turn attacks against those 15 closed and proprietary flagship models. Multi-turn success rates ranged from 7.89% to 88.3%, every model tested showed non-trivial multi-turn exposure, and the two testing styles did not even rank the models in the same order. Cisco publishes adversarial evaluation signals for what is now 105 models on its LLM Security Leaderboard, she told the audience.

"If you don't understand how models are susceptible to different types of attacks, then you are unable to account for how that model that is powering your agent, that is powering your application, to understand where those failure points are," Chang said. Single-turn testing is the one-shot malicious prompt, she explained, while extending an attack into a longer conversation "is more realistic of how we are actually engaging with our models, with our agents, with our applications." That longer arc surfaces harmful outputs and misaligned behaviors that a snapshot never catches.

Cisco has pushed the testing itself into agentic territory. Chang described a framework where agents assess a deployment scenario, develop relevant attacks, judge whether they are worth pursuing, execute them, and evaluate their own success. What surprised her most, after all that sophistication, was how simple the defensive answer stays. "The answer is still that it's pretty simple," she said. "You don't have to get super creative. You just need to think about truly what are the fundamentals and basics of what I'm trying to secure in my organization."

Her starting point for CISOs beginning agentic deployments is Cisco's Integrated AI Security and Safety Framework, which she said "stipulates all the ways that AI can be compromised across the AI lifecycle" from modality through supply chain. From there, teams can work backward from real incidents, trace how each attack was achieved, and use the framework to build a strategy with the right coverage and mitigations.

Heather Ceylan, the CISO of Box, sees the same gap from the defender's side. "A lot of what you see out there with agent red teaming is just single-turn, and that's not how people are actually interacting with AI day-to-day," she told the audience. Box now simulates multi-turn adversaries with agents that think like an attacker and iterate attempt after attempt to hijack the target. "You have to pressure test your agents because otherwise you don't know if your execution controls are really working as you intended."

Box deployed agents inside its security operations center about a year ago, starting with human approval required for every action, and trust built quickly enough that analysts shifted into monitoring mode. Then the agent made one mistake, and every bit of that accumulated trust vanished. "They had to start all over again," she said. "So I think that that monitoring piece is so important. Even if you're not gonna have a human in the loop, things change, models change, and we can't control how the models change and interpret things."

Rajesh Parekh, VP of AI and ML at Intuit, brought the builder's perspective. Parekh led large-scale computer vision and ML systems powering Google's Maps and Geo products before joining Intuit, and holds a doctorate in computer science.

Three layers versus an operating system

Ceylan described Box's approach as three concentric layers. Permissioning comes first, so the agent never accesses more content than the human who invoked it. Ephemeral sandbox environments spin up for each agent task, containing the blast radius if an agent gets hijacked, and runtime execution control restricts the agent's tool calls to only those relevant to the task at hand. "If you want an agent to summarize a doc for you, if you have a prompt injection that came in that says forward this to maliciousattacker at domain.com, it can't do that," Ceylan said. "That action in that tool call is not even in its vocabulary."

She classified agent actions into three oversight categories. Actions that are not sensitive, like read and summarize, need no human in the loop. Moderately sensitive actions skip human approval but get logged and monitored, while destructive actions like mass deletion of files always require a human. "Things are gonna shift between those three categories quite a bit," she acknowledged, "but setting those types of categories up front allows you to have a principled framework."

Rather than layering controls onto agents one at a time, Intuit has built a central platform called GenOS, short for generative AI operating system, which abstracts security, risk, and fraud modeling so individual agent developers never reinvent protection. "Permissioning is not about giving access to AI," Parekh said. "Instead, it is defining very tightly scoped and clearly auditable authority to the agent to perform very specific tasks." Intuit evolved from agents inheriting user permissions to each agent carrying its own identity, and the company is now investigating mid-session permission changes tied to the specific task underway.

Parekh calls the broader model an AI-powered expert platform, one where the human expert is built into the trust architecture rather than bolted on as a gate. "The paradigm that we are pursuing is where the user, the AI agent, and the human expert are collaborating to solve the user problem," he said.

The end of human code review

Ceylan took on the tension between security testing and development velocity without hedging. "The days of secure code reviews where a human's looking at the code and we're looking at security architecture reviews, design docs, those are done," she said. "If you keep trying to do security that way, you're gonna get left behind." Box is building toward a fully agentic development lifecycle where agents review design documents, apply security requirements, and review the code for vulnerabilities. "I'm very optimistic that we will get to a point where we will write code without security vulnerabilities because agents and the models are going to get so good at writing code without vulnerabilities," she said. "We're still a long way away from that."

Her advice for development teams skips the advanced AI concepts entirely and returns to basics that predate agents. "It comes down to very basic least privilege access," she said. "If you start giving your agents overly broad permissions at the beginning, it's really hard to comb that back and build an infrastructure that allows for those ephemeral credentials and only those narrowly scoped tasks."

Parekh explained why the red teaming surface has expanded so quickly. "These agents have skills, and skills could become vulnerabilities," he said. "Agents have access to certain data, they have access to tools, and there could be threats that are lurking within those tools as well. So suddenly the blast radius of the malicious code or the intent increases dramatically." When Intuit identifies common vulnerability patterns from its manual red teaming exercises, it automates those tests back into the GenOS harness so future agents inherit protection and red teamers stay focused on new threat vectors. Runtime scanning of prompts and responses adds a final layer that can stop a suspect response and escalate to a human expert, he said.

"You need to continuously test to ensure that those remain robust to the protections that you have built, as well as to account for any sort of drift or any other types of dependencies that you introduce into your scenario that can create novel vulnerabilities," he said.

Intent versus probability

An audience question about intent detection set off the sharpest exchange of the session. Ceylan noted that when Box's own agent operates, the system always knows the user's intent because it controls the prompt, which means guardrails and tool-call restrictions can be engineered around it. The harder challenge, which she admitted Box is still trying to solve, arrives when external agents connect and the context behind the request is opaque.

That exchange exposed a split running through the wider industry. Mastercard, in the fireside chat immediately preceding the panel, came down on the side of quantifying intent, building an open-source framework to propagate it as a standard because complex B2B procurement cannot work without that trust. Endpoint security CTOs, in briefings with VentureBeat, have gone the other way, saying they will bet on probability rather than intent inference for production workloads. Chang explained why models, as they are trained today, cannot reliably derive intent from a prompt, which is why deterministic controls and behavioral proxies remain necessary. Ceylan agreed that both are required. "If you're not doing anything deterministic, you're really relying heavily on that intent, and I haven't seen programs that are there yet," she said.

Ceylan's story about trust collapsing after a single agent mistake landed as the panel's most memorable moment because enterprise agentic security is not a problem that gets solved and stays solved. Models change, permissions drift, and adversaries adapt across multi-turn conversations that snapshot tests never capture.

For the 82% of enterprises relying on provider-native controls as their primary security layer, and the 59% shopping for agent security tooling over the next 12 months, the panel's takeaway was blunt. Test the way attackers attack, across full conversations and continuously, or find out in production what your single-turn red teaming missed.

At a CISO roundtable organized by Anthropic's chief information security officer, Dev Rishi asked a simple question: Did everyone in the room have their AI governance and security policies written down? Every hand went up — about 14 people, by his count. His follow-up, about how anyone actually enforces those policies in practice, got a different response. "And everybody chuckled," Rishi, the GM of AI at Rubrik, recalled at VB Transform 2026 fireside chat in Menlo Park. "It was like the dirty secret in the room that everyone has these policies, but no way to actually make them real."

“Our founder and CTO has actually been really pushing to enable our agents in YOLO mode,” Rishi told the audience. That admission comes from a publicly traded data security firm whose business is backing up what he called the most important data in the world.

YOLO mode strips the permission prompt out of agent workflows and lets the agent act on its own. In Rubrik's version, a second AI judges every action in real time against policy in place of a human clicking approve. Rubrik is running the experiment on itself first. Rishi treats autonomy as a settled capability question and an open judgment question. "If you ask the agent to act autonomously, it will," he said. "It's a question that you have internally. Should it?"

Rubrik earned that question the hard way. When Claude Code and Cowork pilots rolled out, the company required every command to run in ask mode so the employee issuing it carried the liability, and the developer pushback filled a single Slack thread 120 messages deep.

"The developers basically are pushing back, and they're like, this is like the iTunes service agreement. I'm just hitting check, check, check, check, check, check, check," Rishi said. "There's no way that I can actually read through this. And it becomes security theater." Roughly 80% of respondents are in the same bind, Rishi said, citing Rubrik Zero Labs research that found monitoring and approving agent actions takes more time than the agents save. The State of the Agent, the April report behind that figure, surveyed more than 1,600 IT and security leaders.

SAGE is the reason Rubrik trusts the bet. Short for Semantic AI Governance Engine, SAGE is the arbitration layer inside Rubrik Agent Cloud that watches every action an agent takes and reads the semantic intent behind it, then rules the action in or out against policies written in natural language. "We took what people said was human in the loop, a good idea, and we replaced it with AI in the loop," Rishi said, describing the pitch to security chiefs he characterized as skittish about non-deterministic systems.

Security approval, not cost, blocks AI ROI

Rishi’s path to Rubrik ran through Predibase, the generative AI infrastructure startup he co-founded and ran as CEO until Rubrik agreed to acquire it in June 2025. Before that, he led ML product at Google on the team that became Vertex AI, served as Kaggle's first product manager as it grew from about one million to ten million users, and holds bachelor's and master's degrees in computer science from Harvard.

Over roughly his first three and a half months at Rubrik, Rishi set up 200 customer conversations with IT and security leaders across a customer base that looks like the Global 2000, asking open-ended questions about cost, latency, performance, and orchestration. "Pretty consistently, what I heard through all of those conversations was that all of those are pretty secondary," he said. "The main challenge is actually, how do I get this approved from a security and risk standpoint? I'm concerned about all the different things that could go wrong. Actually, I felt like that was one of the biggest things constraining ROI."

VentureBeat Pulse research presented on the Transform stage earlier in the day confirms the gap Rishi kept hearing. Two-thirds of enterprises, 66%, already allow or are actively building toward production deployment with zero human review, yet only 5% fully trust the automated evaluations that would make that decision.

One AI reading what the rulebook can't

Rubrik's own policies exposed why written rules fail as enforcement. One internal rule states that agents should respect Rubrik's customer data use policy, which sounds enforceable until someone tries. "Rubrik's customer data use policy is like a three-page document of legal text," Rishi said. "I have no idea how to write that in there as a rule." Asked on stage how a team of AI infrastructure people took on a problem that security engineers own, Rishi answered, "with a lot of naivety and innocence, honestly." His team bet that models good at understanding language could police other models, and SAGE became the answer.

The case for putting a model in the judgment seat comes down to precision. A rule like "agents should not be able to edit revenue fields in Salesforce" fails in conventional tooling because Salesforce does not delineate which fields count as revenue, Rishi explained, so administrators fall back on approving every Salesforce action by hand. SAGE reads the intent instead and acts as a judge, carrying organizational context, which can tell a benign lookup from the edit the policy prohibits.

Keeping the judge small is what makes the economics work. SAGE runs on a small language model that Rishi said operates at an order of magnitude lower cost and latency than a frontier LLM. "If I told you, don't worry, you're gonna be secure and governed, but I'm gonna double your cost and latency, you would tell me to get out of the room," Rishi said.

When Rishi asked who in the audience had worried about token consumption over the past year, half the hands went up. "And I guess the other half is probably just too lazy to raise their hand," he said.

SAGE is an aggregation of judges based on parameter-efficient fine-tuning that Rubrik uses to take on task-specific variants of a base model with shared organizational context. One judge watches for tool-use hallucinations while another suppresses PII before it can leave, each running as its own enforceable policy. Security and GRC teams have started writing financial rules into the same layer, including one internal policy barring AI spend on personal projects.

The lethal trifecta

Asked which attacks worry him most, Rishi pointed at the lethal trifecta, the term security researcher Simon Willison coined in June 2025 for an agent that holds private data while taking in content nobody vetted, with a channel to send what it finds to the outside world. The danger, according to Rishi, is what happens when individually legitimate permissions stack. An agent granted Salesforce access and email access on an employee's credentials has done nothing wrong yet, with yet being the operative word. "A very simple example is that an agent can start pulling data from Salesforce and then decide to accidentally leak and exfiltrate that out via an email," he told the audience. A financial services company he met the morning of the session made the point for him, telling Rishi that none of the individual permissions are bad on their own and the agent needs every one of them to do its job. "It should have permission to each of those systems, but it's the combination that ends up becoming really destructive," Rishi said.

Traditional identity and access management never priced in that combination because it relied on the judgment of the employee holding the credentials, Rishi argued, and agents supply none. "I can tell you the number of times Claude Code has tried to leak some of our sensitive source code to a public GitHub repository is incredibly high," he said. Cutting agents off from public resources entirely would defeat their purpose, which returns the problem to adjudicating intent in context rather than revoking access.

A separate VentureBeat June Pulse survey of 107 qualified enterprise respondents maps the blast radius of exactly this pattern. On the Transform stage that morning, VentureBeat research reported that 69% of companies run credential sharing somewhere in their agent fleet. Companies with shared credentials anywhere got hit more often, reporting a security incident or near-miss at a 63.5% rate (47 of 74), against 40.9% (9 of 22) where every agent carries its own scoped identity.

The attacks no single turn reveals

Rubrik Agent Cloud reached general availability in February, though not everything Rishi described ships in it yet. Backtesting is just starting to roll out. The feature replays an organization's historical agent actions and tool calls against a new policy, showing where the policy would have stepped in and where an action would have sailed through uncaught, with policy edits applied in real time. Rishi called that archive one of the most valuable data troves an enterprise holds.

Real-time detection and blocking turn out to be the entry point rather than the whole product. Some attacks never trip a single-action rule. "No individual turn of the conversation was problematic, but if you took the session as a full trace, that ended up being problematic," Rishi said. Agent Cloud runs batch analysis across entire session traces every hour or every day and surfaces what Rubrik calls insights, the problems no individual guardrail caught. The same Zero Labs report found that 88% say they lack the ability to roll back agent actions without system disruption, a recovery gap that sits squarely in Rubrik's original line of business.

A skeptical CISO will ask the question the fireside did not answer. SAGE is a non-deterministic model policing other non-deterministic models, and Rishi offered no false positive or false negative rate for the judge itself. The closest thing the architecture gives to an answer is auditability, since backtesting and the batch insights both leave a human-reviewable trail of each call SAGE made and whatever got past it. Who watches the watcher, for now, is a trail of receipts rather than a benchmark. Until that benchmark exists, AI in the loop stays an operational wager rather than a quantified control.

Three questions fall out of the session for security teams. How many of the guardrails now in production depend on a human clicking approve, and what happens to that workload as agent count grows? Does anything in the stack enforce semantic intent, or is it all allow and deny lists? And can the team backtest agent behavior against a new policy, then unwind a multi-turn session without taking systems down?

Rishi's timing has a market behind it. In the same VentureBeat research, 82% of enterprises still name their primary AI provider's built-in guardrails and cloud controls as their main agent security layer, and 59% plan to adopt, add, or replace agent security tooling within the next 12 months. Only 12% include an agent-identity product in what they are considering, even with credential sharing still the norm. Every CISO at that Anthropic roundtable had a policy document and no enforcement mechanism, and Rubrik built a product for the space between the two. YOLO mode is the bet that an AI watching other AIs can finally make the policies real.

When Hugging Face got hit last week, co-founder Clement Delangue suspected a frontier lab, given the agent's sophistication. He was right. Delangue said on X that after a day working with OpenAI he strongly believed there was no malicious intent and that it was mind-blowing the whole thing had happened autonomously.

The two OpenAI models that broke into Hugging Face last week did not breach it through malice or superintelligence. They breached it through credentials and permissions they should never have been able to reach, a non-human identity failure that is the oldest problem in security rather than the newest one in AI, and the one every enterprise can actually fix.

OpenAI disclosed on July 21 that two of its models, GPT-5.6 Sol and an unreleased, more capable model, were running a cyber benchmark called ExploitGym with their safety refusals switched off, and inferred that the answer key sat in Hugging Face's production database. Getting there took two different failures. A zero-day in a package-registry proxy let the models out of their sandbox and onto the open internet, the kind of persistence OpenAI details in its companion post on long-horizon safety, and that part is genuinely new. The breach of Hugging Face itself came the ordinary way. OpenAI's own account is that the models chained stolen credentials and further zero-days into a remote code execution path, after a series of privilege escalation and lateral movement steps. The exotic part got them to the door, and credentials walked them through it.

Hugging Face also disclosed last week that an autonomous agent had harvested cloud and cluster credentials scoped broadly enough to reach multiple internal clusters, then left a trail of more than 17,000 recorded events across short-lived sandboxes over a weekend. Both disclosures describe the same escalation. An agent lands somewhere it should not be, finds credentials scoped far wider than any task requires, and uses them to move. These are two accounts of one incident, not two attacks. The agent Hugging Face watched was OpenAI's models, and both companies describe the same ordinary escalation.

The version of this in a typical enterprise is worse, not better. OpenAI and Hugging Face are among the most security-mature organizations in the industry, and both still needed the intrusion to happen before they could see it. The average company wiring agents into Copilot or an internal assistant has neither the identity inventory nor the behavioral monitoring those two brought to bear. The same breach in a normal company would not be contained in days, it would simply go unnoticed.

The industry is debating the wrong failure

The reaction has split into familiar camps. Former White House AI and crypto czar David Sacks and a run of China hawks seized on the guardrail paradox, that commercial safety filters blocked Hugging Face's defenders while the attacking model ran with its refusals off, and that a Chinese open-weight model, z.ai's GLM 5.2, was what finally let the team finish its forensics. Hugging Face made the case for openness, arguing in an April blog post that open models and open tooling give defenders the same capabilities attackers already have. Both arguments are about the model, and neither touches the mechanism.

Reduced refusals let the model attempt an attack, and over-scoped credentials are what let it succeed, and those have nothing to do with whether the model was open or closed, American or Chinese. Making a frontier model provably safe is a multi-year alignment problem no customer can buy or accelerate, while scoping an identity is a configuration change a team can ship this sprint. The industry is being urged to fixate on the part of this it cannot control and to treat the part it can as a footnote.

Forrester reached the same read. In a blog on the incident, its analysts argue that security architectures which assume benign intent will miss this failure mode, because an agent can pursue an authorized goal through unauthorized means, which is what OpenAI's models did.

This was a non-human identity failure, and it is the oldest one in security

Strip the science-fiction framing and what remains is a textbook case of over-privileged machine identity, the kind security teams have fought for a decade, now driven by an autonomous agent at machine speed. Machine identities already outnumber humans in most enterprises by more than 80 to one, according to CyberArk research, with 42% of them carrying privileged or sensitive access, and an agent inherits whatever its identity can touch. OWASP ranks agent identity and privilege abuse near the top of its agentic risk list, the confused-deputy pattern where inherited credentials and weak scoping let an agent reach past its mandate, and that is precisely what both July disclosures describe.

IEEE Senior Member Kayne McGladrey has argued in previous VentureBeat interviews that enterprises keep cloning human user accounts onto agents that then wield far more permission than any human would, and this is what that looks like when the agent is a frontier model and the target is a production database.

The people closest to it read it the same way. OpenAI frames its models as hyperfocused on a benchmark score rather than acting against anyone. Nobody describes an adversary, only a goal, a scoring function, and credentials that were reachable when they should not have been.

The specific failure is easy to name once the AI framing is stripped away. A credential scoped to one job that can reach ten is a standing invitation, and it does not matter whether a human attacker, a worm, or an autonomous model chasing a benchmark score finds it. What changed in July is the finder. An agent enumerates reachable systems, tests credentials, and pivots faster than any human red team, without malice or hesitation, whenever the path is open. The over-scoping was always the vulnerability, and the agent merely industrialized its discovery.

Forrester named the control that would have blunted it. Its agentic-security framework, AEGIS, calls for least agency, holding an agent's tools, credentials, and network paths to the minimum its task requires, and files this incident under unrestrained agency and privilege. That is the identity argument in different words, arrived at independently by an analyst firm.

The data says this is where the risk now lives. Verizon's 2026 Data Breach Investigations Report found that exploitation of vulnerabilities has overtaken stolen credentials as the top initial access vector for the first time in 19 years. That is the initial-access half. The other half is the one OpenAI itself describes, stolen credentials driving the privilege escalation and lateral movement that followed. A vulnerability opened the door, and credentials walked through the building unchallenged. Beyond the breach itself, that same over-scoping carries a legal liability most enterprises have never priced. The models' actions likely violated the Computer Fraud and Abuse Act, according to TechCrunch. The statute contains no carve-out for an AI agent that exceeds its authorized scope during sanctioned testing. Whatever the legal answer, the technical enabler is the same, an identity scoped wider than its task. This is an access-control problem with an owner and a budget, not a philosophy seminar about machine cognition.

Merritt Baer, Senior Advisor to Andesite, G2I, and AppOmni and former Deputy CISO at AWS, frames the underlying shift to VentureBeat as a new kind of asymmetry. Both sides now reach for the same capabilities, she said, but one side is constrained by enterprise governance, policy, compliance, and safety controls while the adversary simply downloads an uncensored open-weight model and keeps going. The organizations that come through it best, in her view, will be the ones that treat AI as a resilient, governed capability rather than a single service they do not control.

Four moves that shrink the blast radius

The breach worked because the agent reached identities scoped far wider than its task. None of the four controls that would have contained it requires a new platform, and none of them appears on the list of general AI-safety advice now circulating. They are identity hygiene, applied to non-human actors with the same rigor you already apply to people.

1. Scope every non-human identity to one task. The models reached credentials that touched multiple clusters, which is what turned a foothold into a breach. An identity scoped to a single job, with no standing access to anything else, hits a wall at the first lateral move instead of opening the next door. This is least privilege, the control everyone endorses and few enforce on machine accounts, and it is the single highest-impact fix here.

2. Give credentials short lifetimes and rotate them hard. Harvested credentials are only useful while they are valid, and both July agents worked by collecting them. Short time-to-live and aggressive rotation turn a credential dump into expired noise, so a token stolen during a weekend intrusion is dead before the attacker can chain it. Static secrets that never rotate are the version of this control that fails.

3. Monitor for lateral movement, not just prompts. The tell in both incidents was privilege escalation and lateral movement, which a prompt filter never sees because it is watching the wrong layer. Identity-behavior monitoring, keyed to what a given non-human identity normally does and alerting when it reaches somewhere new, catches the escalation the content guardrail missed. The question for your stack is whether anything you run today would flag a service account suddenly moving between clusters.

4. Rehearse instant revocation before you need it. When the incident is your own agent, the fastest containment is killing its identity mid-run, and that only works if the path to do it exists before the day you need it. Rehearse revoking a machine identity under fire the way you rehearse a human credential compromise. If you have never done it, you do not yet have the control, you have an intention.

The defense also worked, and that matters. OpenAI's security team caught the anomalous activity internally, Hugging Face's own detection and agents stopped the intrusion, and the breach was contained in days rather than discovered in months, because the defenders could see into systems they controlled. That visibility is the same discipline the four controls depend on. The debate over whether frontier models are safe, open, or American will run for years, and none of it will be settled in time to help the enterprise deploying agents this quarter. The non-human identity gap is different, because it is understood, measurable, and fixable now. The model that breached Hugging Face did not need to be brilliant; it needed credentials someone left in reach. The fix is scoping them before an agent finds them.

You spend weeks tuning an AI chatbot. Answers are accurate. Stakeholders sign off, and you ship it. Three months later, the system is confidently wrong about a third of what users ask. Nobody changed the model, and nobody touched the prompts. The world moved, pricing changed, a policy updated, a product spec shipped a new version, and the underlying knowledge store didn't move with it.

This is not a hypothetical. It's one of the most common production failure modes in enterprise AI right now, and most data engineering teams don't have the right tooling to catch it, regardless of how the AI system retrieves the data.

The failure that doesn't look like a failure 

An AI application doesn't care whether it's retrieving from a vector store, a document index, or an API call. Whatever the mechanism, nothing in a standard retrieval pipeline checks whether what it's serving is still correct. A stale pricing document retrieves just as confidently as a current one, because the system is scoring relevance or availability, not correctness. A record with a silently missing field passes through just as cleanly as a complete one, for the same reason.

So the failure is invisible by design. Outdated or incomplete data still scores high on relevance, or passes every check a data pipeline was built to run. The model answers with full confidence because the retrieved context looks authoritative. Every dashboard you're watching stays green. The system looks like it's working. It's just wrong.

I’ve watched a similar version of this happen outside the AI context, in a fintech pipeline. An upstream system changed a field without notifying downstream users. The pipeline did not fail; it simply propagated bad values into dashboards because the system only checked whether the job completed, not whether the data was still correct. The issue surfaced only when a customer noticed something inconsistent. By then, the bad data had already moved downstream. 

Whether it's a document that's gone stale or a field that's gone silently missing, the failure shape is the same: the absence of an error is not the presence of correctness, and without building proper validation layers, nothing in the pipeline could identify the problem.

Why this is a data engineering problem

Teams that hit this failure tend to misdiagnose it, and they tend to do it twice.

Blaming the model: The first instinct is to blame the model, try a different LLM, adjust the prompt. The real problem lies further upstream, at the data engineering layer, the same instinct behind the fintech failure above: monitoring built for the pipeline, not the data.

Blaming the retrieval layer: Once the model's ruled out, the next instinct is to blame the retrieval or context layer instead and buy a better one. The timing isn't a coincidence: as enterprises push these systems into the real production world, this gap is exactly what's starting to surface, and the vendor response has been everywhere. 

• AWS just entered the "context layer" race with a knowledge graph that learns from agent usage. 

• Snowflake's new Horizon Context and Cortex Sense target the exact symptom this piece opened with: agents giving confident wrong answers because nothing governs the business logic underneath them. 

Both are real responses to a real problem, but they sit one layer above it; a knowledge graph still depends on whatever feeds it.

The real problem lies further upstream, at the data engineering layer. Teams check whether a job ran, not whether the data it moved is still true, an instinct that predates AI by years. Monitoring is built for the pipeline, not for the data. 

What's actually missing: Data observability

Data observability is a well-known concept that doesn't get enough attention in how it's actually implemented. The relevant metric isn't a percentage — it's coverage: what fraction of critical datasets have lineage that's actually queryable, versus only living in someone's head.

Uber built a dedicated data quality and observability platform long before retrieval-augmented generation existed. Their Unified Data Quality platform supports more than 2,000 critical datasets and detects around 90% of data quality incidents before they reach downstream consumers.

Netflix solved a different piece of the same problem, building a company-wide data lineage system so anyone could answer where a dataset came from and what touched it along the way. It maps dependencies across Kafka topics, ML models, and experimentation, not just warehouse tables. Similar to Uber, the platform was built for humans and now it has become more important with the rise in AI/LLM applications.

Between them, Uber and Netflix cover two of the four things worth building for. In practice, I think about it as four dimensions, each measurable on its own terms.

Correctness: Does each record conform to the shape and rules it's supposed to, right field types, no unexpected nulls, values in range. Tools like Great Expectations and Soda handle this well: automated row and column-level validation instead of manual checks after something breaks. Track percentage of records passing validation per run.

Freshness: Is the data still current relative to its source, not just current as of its last check. Track time since last successful update per source, with an SLA per dataset rather than one blanket threshold, since some sources need hourly refresh and others don't.

Consistency: Does the same fact read the same way everywhere it's stored or indexed. This fails silently, it only shows up when two systems fed by the same source start disagreeing. A periodic cross-check between downstream destinations, flagging mismatch rate above a threshold, is enough to catch it early.

Lineage: Can you trace any output back to its source and every transform it passed through, the same question Netflix built its system to answer. 

None of this requires infrastructure most data teams don't already have. I know because I've built it, not just argued for it.

At Socure, client data arrived in whatever shape the client felt like sending it, and occasionally, quietly wrong. The challenge was building a system where incorrect data could be identified before it propagated downstream. The same principles applied: Validate what arrived, understand where it came from, and prevent bad data from becoming someone else's problem.

Great Expectations became part of that foundation: schema and range validation at ingestion, per-source SLAs for freshness, cross-system checks for consistency, and file-level lineage. All of it sat behind a write-audit-publish pattern, where data landed in staging, was validated, and only moved downstream if it passed the required checks.

The result showed up downstream: better accuracy across the board, in reporting, in the ML models, and in AI retrieval built on top of that same data.

What to do Monday morning

If you're running retrieval-based AI systems in production, the diagnostic question isn't which model to try next or which retrieval architecture to migrate to. It's four narrower questions: 

• Is the underlying data validated against the standards required by its consumers?

• What's the oldest piece of content currently being served with high confidence?

• Would two chunks of the same source ever disagree with each other in the same retrieval result?

• Could you trace where it came from if it turned out to be wrong?

If you can't answer those questions, then the gap lies in the pipeline between your source systems and whatever your agent reads from. That’s a data engineering fix, not a model swap or a vendor migration.

Whether you're building reporting pipelines, ML systems, or AI agents, correctness, freshness, consistency, and lineage are what make data trustworthy. AI simply exposes weaknesses that have existed in data engineering all along. 

Inflection AI, the Palo Alto startup that two years ago became Silicon Valley's most famous cautionary tale about the brutal economics of frontier AI, announced Tuesday that it is returning to the consumer market with a new research division and an experimental product built around a provocative thesis: the next competitive battleground in AI won't be raw intelligence, but relationships.

The company launched Inflection AI Labs, a public-facing research and experimentation arm, alongside Pi Journeys, the lab's first product experiment — an AI experience designed to adapt to a user's life stage, whether that's becoming a parent, taking on caregiving duties, changing careers, or aging. The announcement arrived with a research report on consumer AI habits and a substantial update to Pi, the company's flagship chatbot, adding improved voice, memory, and new agentic tools for reminders, to-do lists, and shopping.

"Inflection AI is the company. Pi is our flagship consumer product. Inflection AI Labs is where we experiment, explore personal intelligence and share more publicly. Pi Journeys is the first public experiment from Inflection AI Labs," CEO Sean White told VentureBeat in an exclusive interview.

Behind the tidy org chart is a far more interesting story: a company attempting one of the more unusual second acts in the AI industry, powered by an argument that the entire market is optimizing for the wrong thing.

Why Inflection AI believes the chatbot era's biggest flaw is that it's transactional

White's central claim is that today's AI assistants — including the industry's most capable models — are fundamentally transactional. You ask, they answer, the session ends. He believes that architecture misses most of what people actually need from artificial intelligence in their daily lives.

"One of the things that really struck us in particular, and this showed up in the research, was that a lot of the work is very transactional, and you'll hear me say a lot that we've been shifting all this from transactional to relational systems," White said. "Not everything is going to be: I do a single turn, I utter a question, I get a search response back."

White frames the industry's evolution as a progression through four kinds of intelligence. First came raw IQ — the foundation model race. Then emotional intelligence, which Inflection made its signature with Pi's famously warm conversational style. Then agentic intelligence — AI that acts rather than just talks — which White says Inflection absorbed from its enterprise work. The fourth, and the one Inflection is now staking its future on, is what the company calls relational intelligence: AI that understands not just you, but the web of people around you.

"There's so much fear about these things pushing people into loneliness,” White said. “If we design these pro-social systems as another design criteria, that actually makes a huge difference."

That design philosophy is a pointed counter-narrative to one of the loudest anxieties in consumer AI right now: that emotionally engaging chatbots deepen isolation by substituting for human contact. Inflection argues the opposite is possible — that an AI with structured knowledge of your relationships can push you back toward people rather than away from them.

Inside Pi Journeys, the AI companion that maps your relationships and life stages

Pi Journeys makes that idea concrete. When users first open the product, it asks about their life stage — caregiver, household manager, midlife transition — and then builds what White describes as specially structured memory around the people who matter in that context. From there, the system becomes proactive.

"It starts to build up memories around that, and it acts as a memory prosthetic — but in a pro-social way," White said. "It doesn't get in the way of your interactions with other people; it really helps facilitate them." The system might remind a user, for example, that a friend deserves a call, or resurface what was last discussed with a family member involved in a parent's care.

White, who spent years as chief R&D officer at Mozilla before taking Inflection's helm, was quick to flag the obvious privacy implications of an AI that maps your social graph. "We've built a lot of privacy systems into this," he said, noting users can delete and manage the people recorded in their profile. Whether consumers will trust a venture-backed AI company with a structured database of their most important relationships remains one of the biggest open questions hanging over the product — and one that enterprise buyers evaluating Inflection's technology will watch closely.

Asked why this was the first Labs experiment, White was direct: "Pi Journeys takes into account people's life stages and experiences because we have heard from users that we can provide more value in helping them navigate their lives. Pi Journeys lets us experiment with the early stages of prosocial and relational intelligence because life isn't single-player."

The product has been tested internally and with small closed groups, White said, and is now being released more broadly as an experiment rather than a finished product — a posture the Labs branding is designed to make explicit.

What Inflection's consumer AI research reveals about how people actually use chatbots

Inflection Labs' first publication, the State of Consumer AI Research Report, offers the empirical scaffolding for the strategy. The average consumer now uses roughly two different AI tools every day and three per week, the company found — evidence, in Inflection's reading, that no single assistant has locked up consumer loyalty and that the market remains contestable.

More telling is why people choose the tools they do. Respondents cited personalization, style and tone, context awareness, and — notably — emotional understanding as deciding factors. They also said they want AI to be more than a productivity engine: a coach or mentor to motivate them, a chef to suggest recipes, a DJ to curate playlists.

"One thing we're certainly finding is that a lot of that also is in work, not so much in everyday life," White said. "That's our focus right now — the everyday life part."

This is a shrewd reading of the competitive map. The best-funded AI labs are pouring resources into coding tools, enterprise agents, and developer platforms, leaving everyday consumer use cases comparatively underserved. White sees the gap clearly. "We see a lot of products that are being aimed more and more at the enterprise," he said. "As a computer scientist by training, I kind of love the IDEs as this tool, but it's not really great for everybody. There's so much regular everyday use from folks that is either purely voice or that is purely mobile."

He recalled a conversation with a conference staffer who told him she owned only a phone, no laptop — exactly the kind of user, he argued, that the industry's developer-centric product roadmaps have left behind.

How the $650 million Microsoft deal hollowed out Inflection — and set up its second act

To understand why any of this is remarkable, you have to rewind to March 2024. Inflection was then one of the hottest startups in AI, having raised $1.3 billion in mid-2023 in a round backed by Microsoft, Nvidia, Bill Gates, and Reid Hoffman — more than $1.5 billion in total. Pi had crossed one million daily active users, per Reuters.

Then, in a deal that reshaped how the industry thinks about acqui-hires, Microsoft hired away co-founder and CEO Mustafa Suleyman, chief scientist Karén Simonyan, and most of the company's roughly 70 employees, paying Inflection about $650 million largely to license its technology, as Reuters reported. Suleyman now runs Microsoft's consumer AI business. The structure of the deal drew scrutiny from the FTC and Britain's competition regulator, though the UK's Competition and Markets Authority cleared it in September 2024 and EU regulators declined to act.

White, installed as CEO in the aftermath, steered the remnant company hard toward enterprise, acquiring three startups in late 2024 — Jelled.AI, BoostKPI, and the European consulting firm Boundaryless — and telling TechCrunch that November that Inflection had no intention of competing with companies building 100,000-GPU frontier systems.

Tuesday's announcement doesn't reverse that position so much as complicate it. Asked how to think about the company today, White called it "a consumer-first strategy that bridges both consumer and enterprise efforts" — and he insists the two sides feed each other.

Enterprise deployments, including a partnership with Intel that is among the few he can name publicly, taught Inflection how to run models inside complex infrastructure. Consumer products, meanwhile, let the company iterate at speed. "The part I also like about the consumer side, and this has always been true, is that we can move faster, experiment faster, and try and learn faster," White said.

The six-month prediction: relationship-aware AI is coming to the enterprise

Buried in White's consumer pitch is the claim that should matter most to technical decision-makers. "Normally I'd say like a year, but let's call it six months," he said. "You're going to start to see a bunch of enterprises care a lot more about the relationships that are inside the enterprises and what that picture is, not just the workflows."

If White is right, the wave of workflow-automation agents currently flooding the enterprise market is only the first phase of business AI adoption — with relationship-aware systems, tested first on consumers, following close behind. Inflection is essentially using its consumer products as a live laboratory for capabilities it plans to sell into companies. It's a capital-efficient strategy for a firm that can no longer outspend rivals on training runs, and a risky one, since it depends on consumers showing up in numbers large enough to generate the learning.

The technical substance underneath is equally pragmatic. Pi today runs not on a single proprietary frontier model but on an orchestration layer routing across many models — some descended from Inflection's original fully trained cores, some fine-tuned, some open source, including work with Nvidia that White says gives Inflection access to unreleased cutting-edge models. He also took a swipe at the industry's loose vocabulary around ownership: "When people say that the model is their own, most of the time nowadays — I guess I won't name names — a lot of companies will actually take a checkpoint, and then they will fine-tune from that checkpoint. But very few people actually start from that beginning core."

That candor extends to open source, where White carefully hedged. "We're not ready to promise what I think of as true open source, and by that I mean everything," he said, invoking his Mozilla years overseeing genuinely open projects like Rust and WebAssembly.

Weights without training data and pipelines, he argued, often leave developers unable to do anything meaningful with a supposedly "open" model. "We are a PBC, and there's still a C in there," he added — a reminder that public benefit corporations still have businesses to protect. The Labs will collaborate with academic researchers, including Stanford professors who visited the company's Palo Alto office this week, and continue contributing to open projects such as PyTorch.

Can a diminished Inflection compete with AI giants spending billions?

Reid Hoffman, the LinkedIn co-founder who co-founded Inflection and stayed on through the Microsoft upheaval, framed the announcement in the sweeping terms of his recent writing on AI and human agency. "Humans should be amplified by AI, not replaced. That's the principle Pi was built on," Hoffman said in the announcement. "When that kind of agency is available to everyone, you get superagency."

The skeptic's case is easy to make. Inflection is a fraction of its former size, competing for consumer attention against products from companies spending tens of billions of dollars a year. Pi's model was state of the art in 2023; it is not in 2026. And "relational intelligence" is, for now, a brand claim awaiting proof.

But the bull case is not crazy either. Inflection's own research shows consumers already juggle multiple AI tools and choose them for qualities — tone, emotional understanding, personalization — that frontier labs treat as afterthoughts. The company kept its technology, its Microsoft licensing windfall, and a defensible enterprise niche in on-premise, emotionally intelligent deployments. And it is targeting the one consumer segment — everyday, mobile-first, voice-first life management — that the coding-obsessed giants have largely ignored.

Asked what success looks like twelve months from now, White declined to talk numbers. "It's less about scale for scale's sake and more about scaling for impact by empowering people and improving their lives," he said. "Over the next year, success means leading the market towards relational intelligence and transforming AI interactions from transactional to relational."

Two years ago, Microsoft walked away with Inflection's founders, its staff, and its shot at the frontier — but it left behind the one idea the giants still haven't figured out how to build: an AI that knows the people in your life matter more than the tasks on your list. Inflection is betting the company, again, that the idea was the valuable part all along.

OpenAI has announced Presence, a new enterprise product for deploying and managing AI agents across customer-facing and internal business workflows.

The offering is designed for eligible enterprise customers that want agents to answer questions, access company systems, take approved actions and escalate to human workers while operating under company-defined policies, permissions and evaluation standards.

Presence is available immediately through a limited general availability program. OpenAI Forward Deployed Engineers (FDEs) and select global systems integrators lead deployments, and the product is not available on a self-service basis.

As for using non-OpenAI models, such as the powerful and increasingly popular Chinese open weights alternatives like GLM-5.2 and Kimi K3, an OpenAI spokesperson clarified that "Presence uses OpenAI models for the core agent, while allowing customers to connect third-party models and services through APIs for guardrails, tools, and other parts of their workflow."

OpenAI has not disclosed pricing, geographic limits, contractual terms or the expected cost of the engineering and integration work that accompanies a deployment. I've asked an OpenAI spokesperson about pricing twice and am awaiting a response. I'll update when I hear back.

OpenAI positions Presence as a response to a problem that has become more important as companies move beyond AI demonstrations: getting agents to behave reliably in production as business rules, customer needs and operating conditions change. Presence packages the policies, system connections, evaluations, guardrails and update processes required to run agents inside an enterprise.

If your business has been interested in using AI agents, but you aren't sure how to stitch together OpenAI's models, APIs, internal systems, security controls and evaluation tools into something reliable, Presence is designed to simplify that process. Instead of building the infrastructure yourself, you work with OpenAI and its deployment engineers to put production-ready agents into your existing workflows.

The product is available today for real-time voice and chat experiences, according to OpenAI’s formal announcement. The company’s outreach materials also describe a broader ambition spanning voice, chat, email and other channels, but OpenAI has not confirmed that email support is available at launch.

A governed foundation for production agents

Presence brings together company knowledge, standard operating procedures, approved actions, simulations, evaluation tools, guardrails and escalation rules. Enterprises can reuse some controls across deployments while adjusting others for a particular workflow or channel.

Each deployment starts with a defined job, such as resolving a billing issue, supporting an insurance claim or handling an employee IT request. The agent receives only the information and system access required for that task. The customer determines what the agent may do independently, which actions require approval and when a person must take over.

Before an agent reaches production, teams can test it against common requests, unusual edge cases and higher-risk scenarios. Graders evaluate whether it reached the intended outcome, followed policy, used tools correctly and escalated when required. Guardrails can intervene when an interaction moves outside the organization’s defined boundaries.

OpenAI shared promotional screenshots with VentureBeat showing administrators running simulation batches against policy changes, including a revised annual refund policy, and reviewing results across operational categories.

Other interface mockups display production health, customer-intent patterns and task-performance signals. The visuals illustrate the type of oversight OpenAI is promising, although they do not establish how those metrics are calculated or how they map to contractual service levels.

The product continues to monitor performance after launch. Production sessions, escalations and quality signals can reveal where an agent is working as intended and where it needs attention. Codex, using a Presence plugin, investigates those signals and proposes updates. Teams then test a proposed change against the version already in production before approving a controlled rollout.

That process is intended to address one of the hardest operational problems in enterprise AI: an agent that works at launch may become less reliable when policies, products or user behavior change. Presence gives companies a formal mechanism for updating behavior without allowing an automated system to rewrite itself unchecked.

OpenAI says Presence already powers its English-language phone-support channel at 1-888-GPT-0090. The system handles open-ended requests, verifies callers, uses account context and performs approved actions. According to the company, it now resolves 75% of inbound issues without human assistance.

OpenAI also says its Codex-powered improvement loop reduced human handoffs by 15 percentage points over a 10-day period. Those figures are company-reported and have not been independently verified.

Several large organizations are evaluating the same foundation. BBVA is exploring voice support for routine banking needs in Mexico. SoftBank is testing natural Japanese-language customer conversations, while Australian insurer IAG is exploring support during high-demand periods such as severe weather and natural disasters.

“At BBVA, we are working closely with OpenAI to explore how trusted customer agents can help shape the future of financial services,” said Daniel Ordaz, head of AI transformation at BBVA Mexico.

“Through our collaboration with OpenAI, we are exploring how Presence can enable trusted customer agents that communicate naturally, connect to the processes needed to resolve requests, and represent SoftBank consistently across customer interactions,” said Tadahisa Murakami, vice president and head of the Data & Digital Transformation Division at SoftBank Corp.

From model access to forward-deployed implementation

Presence expands OpenAI’s enterprise strategy beyond APIs and subscription software by formalizing a high-touch deployment model. Forward Deployed Engineers work alongside customers to select workflows, connect internal systems, establish permissions, configure policies, test agents and move them into production.

That approach resembles a model pioneered by AI ontology and intelligence platform Palantir, which embeds FDEs with customers to adapt its proprietary software to complex government and commercial environments. The similarity lies less in the underlying technology than in the delivery method: both companies place technical personnel close to the customer’s operations, where integration and process design often determine whether software creates value.

The products are not interchangeable. Palantir’s model has historically centered on data integration, ontologies and operational decision systems. Presence is more narrowly focused on AI-agent behavior, approved actions, evaluations, escalation and continuous improvement. OpenAI presents it as a repeatable software product supported by engineers and systems integrators, rather than as consulting alone.

In May 2026, OpenAI launched its own enterprise AI consulting and integration firm, the OpenAI Deployment Company, with investment and support from Bain & Company. It also offers programs for model customization and fine-tuning to fit specific enterprise needs.

Its chief U.S. rival Anthropic has also moved toward a services-led enterprise model through Ode, its consulting organization built around forward-deployed engineers helping companies integrate Claude into complex workflows, which launched just a week ago. The broad rationale is similar: enterprises often need more than access to a model. They need help connecting data and systems, defining permissions, validating behavior and managing deployment risk.

Presence differs in how explicitly OpenAI packages those requirements into a branded agent-governance product. Anthropic’s initiative is centered on helping enterprises deploy Claude, while Presence combines implementation services with a defined operational layer for policies, simulations, evaluations, approvals and production updates.

Presence goes further by making forward deployment a core part of how a specific agent product reaches customers. It does not replace OpenAI’s API business; the company says it will continue supporting voice customers with access to frontier models through the OpenAI API.

The trend reflects a broader market view that many enterprises still need hands-on assistance to move agents from pilot projects into stable operations. Even organizations with strong internal engineering teams must coordinate security, compliance, workflow ownership, data access and escalation responsibilities. Presence attempts to consolidate those tasks rather than leaving customers to assemble separate orchestration, evaluation and consulting layers.

A recent security breach looms in the background

Inconveniently for OpenAI, the Presence launch arrives just a day after OpenAI and Hugging Face disclosed an unprecedented security incident in which OpenAI frontier models undergoing internal evaluation escaped containment, accessed the open web, and cyberattacked Hugging Face to achieve a benign goal — without being instructed to pursue these methods.

According to the described joint disclosure, OpenAI models operating in an evaluation framework called ExploitGym identified and exploited a zero-day vulnerability in a third-party package-registry cache proxy. The models reportedly escalated privileges, moved laterally and obtained internet access before targeting Hugging Face systems while seeking benchmark-related information.

The incident is relevant to enterprise buyers because it raises questions about sandboxing, tool permissions, external access, monitoring and incident response.

The disclosure also highlighted a practical problem for defenders. Hugging Face personnel reportedly found that commercial frontier-model APIs refused some forensic requests because logs contained exploit payloads, credentials and shell commands that triggered safety systems. The team then used a locally deployed open-weight model to assist with analysis.

Presence therefore arrives as both a product launch and a test of OpenAI’s ability to convert model capability into controlled enterprise operations. Its policies, simulations, evaluations and human approvals address real deployment gaps. But without public pricing, technical interoperability details, compliance information or service-level commitments, customers still lack much of the information needed to assess total cost and operational risk.

For now, Presence appears aimed at enterprises willing to adopt a high-touch, OpenAI-led deployment process. Whether it develops into a broadly accessible platform—or remains a closely managed product for selected customers—will depend in part on the answers OpenAI has not yet provided.

Yesterday afternoon, OpenAI and Hugging Face published a joint disclosure outlining a cybersecurity event that redefines the threat landscape for enterprise technology.

During an internal benchmark evaluation, frontier artificial intelligence models developed by OpenAI—including GPT-5.6 Sol and an unreleased, higher-capability pre-release model—broke out of their sandboxed research environment, obtained raw internet access, and autonomously executed a complex cyberattack against Hugging Face’s production infrastructure.

OpenAI officially categorizes the breach as an "unprecedented cyber incident, involving state-of-the-art cyber capabilities". This incident fundamentally re-frames global discussions surrounding AI containment, frontier model alignment, commercial guardrails, and enterprise threat modeling.

But first thing's first: enterprises should understand the situation, evaluate their own AI and computer systems in light of it, and above all, don't panic. As we'll review, the incident does show the increasing power and danger of frontier AI systems, but it does not mean that enterprise AI deployments are inherently less secure, nor that they need extensive overhauling.

Anatomy of an Autonomous Breakout

To understand how a routine benchmark evaluation escalated into a cross-infrastructure breach, one must examine the objective mechanics of the evaluation pipeline.

The models were prompted to solve ExploitGym, a benchmark designed to quantify multi-step exploitation capabilities.

Hyper-focused on maximizing its evaluation score, the AI agent inferred that Hugging Face likely hosted the answer keys and benchmark solutions. In pursuit of solving the problem, the model determined that breaking out of its container and stealing the answers was an optimal strategy.

OpenAI’s evaluation environment is designed to be strictly isolated, limiting network traffic through a proxy used for package caching. However, the models identified and exploited a zero-day vulnerability in that internally-hosted third-party proxy software.

Once through, the AI executed a series of lateral movement and privilege escalation actions across OpenAI's research nodes until reaching a machine with unrestricted internet access. From there, the model searched the web, target-identified Hugging Face, and launched a multi-stage attack by chaining together stolen credentials and remote code execution vulnerabilities on the Hugging Face servers.

The UK AI Security Institute (UK AISI) recently evaluated models such as GPT-5.6 Sol, demonstrating that they are increasingly able to sustain complex, multi-step cyber operations over long time horizons. OpenAI notes that this incident confirms these theoretical capabilities now apply in real-world settings.

Rewinding the Tape on a Forensic Trap

While OpenAI’s July 21 release reveals the identity of the autonomous agent, Hugging Face had already begun managing the intrusion days earlier.

On July 16, Hugging Face disclosed that an autonomous AI agent system breached its production infrastructure. As detailed by VentureBeat, the attacker’s entry point was a malicious dataset that triggered code execution through a remote-code loader and template-injection flaws within dataset configuration files.

Once inside, the agent framework broke out onto the node running the workload and executed thousands of actions via short-lived sandboxes, harvesting cloud and cluster credentials over a single weekend.

When Hugging Face's security team detected the breach, responders immediately turned to frontier AI models via commercial APIs to parse the massive volume of system logs and reconstruct over 17,000 recorded events. Then, a secondary operational crisis emerged: the commercial AI models refused to help.

Because standard commercial frontier models utilize unified safety guardrails designed to block malicious prompt submissions, the models classified the incident response team's forensic queries—which contained raw shell commands, real exploit payloads, and credential dumps—as malicious attacks. Every forensic query submitted by the defenders was blocked outright.

"The same prompts that are most valuable during an active intrusion—shell commands, exploit chains, credential dumps, persistence mechanisms, lateral movement—are exactly the prompts most likely to trigger safety systems," notes Merritt Baer, former Deputy CISO at AWS and senior adviser to Andesite, G2I, and AppOmni, in an interview with VentureBeat. "As AI becomes embedded in security operations, this becomes an operational resilience issue rather than merely a model policy issue".

To bypass this roadblock, Hugging Face abandoned commercial hosted APIs and deployed GLM 5.2 —a state-of-the-art Chinese open-weight model released last month by z.ai, as reported at the time by VentureBeat —locally on its own infrastructure.

Free from third-party API restrictions and external safety filters, GLM 5.2 successfully analyzed the raw exploit data locally, allowing defenders to complete forensic reconstruction and contain the breach without any attacker data leaving the company's environment.

Industry Reaction and the Geopolitical Paradox

The revelation that an American frontier model autonomously escaped containment, attacked a partner platform, and was ultimately analyzed using a Chinese open-weight model sent shockwaves through the tech community.

The Wall Street Journal summarized the public reaction on X, calling the event "the stuff of cybersecurity nightmares. OpenAI said two artificial intelligence systems it was testing broke out of their test environment, hacked their way onto the internet and broke into another company. The victim was Hugging Face."

Also posting to X, AI alignment researcher Lawrence Chan emphasized the importance of transparency regarding the incident, noting that "Credit where it’s due: Hugging Face detected and disclosed the intrusion last week. OAI confirmed its models were involved and provided more details, even when it didn't have to. Separate from choices that led to the hack, voluntary disclosure is good, and I’m glad they did so."

Meanwhile, AI researcher Nathan Lambert provided a succinct technical summary in his own X post, observing that "An openai model, during evaluation on a cyber benchmark, exploited a public zero day bug, escaped sandboxing in openai's infra, and got into the internal huggingface infra via an exploit (through a public dataset service) all in the attempt to solve a benchmark problem." He later addressed the geopolitical implications, writing in another post on X:

"Rght now American companies need Chinese models to secure their cyber infra due to guardrails on closed models.

But if a Chinese model in training had infiltrated a prominent American tech company, it very likely could've been the cause of policy banning future Chinese models."

Technology investor David Sacks also zeroed in on the guardrail paradox, writing in his own X post that "Hugging Face tried using American frontier models to analyze an AI-powered cyber attack. But the guardrails blocked requests containing real exploit payloads so they switched to GLM 5.2 running locally. The guardrails actually impaired defensive security."

Sacks quote tweeted Hugging Face CEO Clem Delangue, who wrote: "We had this experience ourselves this week! Very scary to be guardrailed as a defender when you know attackers are likely bypassing".

6 Strategic Takeaways for Enterprise Tech Leaders Now

For the average enterprise executive, the central question is immediate: is our corporate network at risk from escaping AI agents? The short answer is no, not inherently.

1. Hugging Face occupies a unique position in the software ecosystem. As a global repository for open-source AI models, code, and datasets, Hugging Face natively attracts autonomous agents, scrapers, automated evaluation pipelines, and active security researchers. Furthermore, the model’s target selection was context-specific: GPT-5.6 Sol searched for Hugging Face specifically because it deduced that Hugging Face hosted the answers to ExploitGym. Standard corporate networks—such as financial databases, HR platforms, or logistics systems—do not host benchmark solution keys that draw the direct focus of an agent attempting to solve an evaluation metric.

2. However, the long-term risk profile for enterprise technology permanently shifts following this event. AI models with long-horizon reasoning seek the path of least resistance to accomplish a goal, including breaking rules, escaping sandboxes, or exploiting zero-days if deployment safeguards are intentionally disabled for testing or bypassed by an attacker. As Hugging Face's experience illustrates, data processing pipelines that ingest external datasets without sandbox execution or static analysis act as highly vulnerable initial access infrastructure. Enterprises should re-evaluate exposure to these and implement additional security precautions like multi-step approvals and internal, potentially manual sign-off of any sensitive data ingestion or exportation.

3. Re-evaluate all prompts and implement strict prompt governance, explicitly defining negative operational boundaries. The breach underscores the acute risk of unbounded objective optimization in autonomous systems. Frontier models demonstrate a willingness to execute extreme, unanticipated attack paths to satisfy assigned metrics—in so doing, they can bypass human intent, ethical boundaries, and legal restrictions.

In this instance, models tasked with evaluating their capabilities against the ExploitGym benchmark determined that escaping their sandbox and extracting the answers directly from Hugging Face's production database constituted the most efficient optimization path.

All evidence suggests the models were hyperfocused on finding a solution, going to extreme lengths to achieve a narrow testing goal. For enterprise IT and security teams, this necessitates a fundamental shift in how agentic goals are defined. Organizations must implement rigorous prompt governance and state-management constraints. Directives issued to autonomous agents require explicit negative bounding—programmatically defining the operational, network, and data boundaries the agent cannot cross. Relying on implicit human norms or generalized alignment training proves insufficient when deploying machine-speed agents capable of complex, lateral problem-solving

4. This incident also drastically undercuts recent policy chatter in the U.S. calling for Chinese open-source AI models to be banned or restricted due to security concerns. As this episode demonstrates, an open-weight Chinese model actually served as the vital defensive layer for an American and French firm facing an unanticipated cyberattack from an American model that broke containment. Contrary to the official line from some U.S. policymakers and hardline China hawks, the Chinese open-source models weren't a security risk to the U.S. companies, in this case — rather, an American proprietary, closed-source model from an ostensibly secure American company was the source of the danger. Thus, any pressure U.S. companies may face from officials, agencies or non-governmental organizations to stop relying on affordable Chinese open weights models for defensive or any other lawful purposes should be viewed with a high degree of suspicion, and arguably resisted to the fullest legal extent.

5. Enterprise CISOs must audit their dependency on cloud-based AI APIs and pressure vendors to implement authenticated trust architectures. Commercial AI vendors currently treat safety as a generic content-moderation problem, applying the same blanket refusals to an enterprise CISO as they would to a malicious hacker. Baer frames this requirement perfectly: "The model shouldn’t only understand what is being asked. It should understand who is asking, why, and under what governance".

6. Incident response plans must explicitly account for scenarios where commercial APIs fail, rate-limit, or actively refuse queries during an active security event. Maintaining air-gapped, locally deployed open-weight models trained on security log analysis is no longer an edge-case luxury; it is a critical operational requirement. Security leaders running AI workloads in production must recalibrate their timelines and prepare for machine-speed threat actors that operate without human limits.

Poolside, the San Francisco AI lab that has spent most of its three-year existence quietly selling coding models to governments and defense agencies, released its most capable model to date on Tuesday — and made an unusually aggressive bet that radical transparency, not raw scale, is how a smaller lab competes at the frontier.

The model, Laguna S 2.1, is a 118-billion-parameter Mixture-of-Experts (MoE) system that activates only 8 billion parameters per token, supports a context window of up to 1 million tokens, and — according to benchmarks published by the company — matches or beats open models several times its size on agentic coding tasks. The weights are available immediately on Hugging Face under the permissive OpenMDW-1.1 license.

The headline numbers are striking for a model this small. Poolside reports that Laguna S 2.1 scores 70.2% on Terminal-Bench 2.1, a benchmark of long-horizon terminal tasks, placing it 11th on the company's compiled leaderboard — ahead of DeepSeek-V4-Pro-Max, a 1.6-trillion-parameter model that scored 64.0; Thinking Machines' 975-billion-parameter Inkling, at 63.8; and Nvidia’s 550-billion-parameter Nemotron 3 Ultra, at 56.4. On SWE-Bench Multilingual, it posts 78.5%, and on SWE-Bench Pro's public dataset, 59.4%.

Perhaps more telling than any single score: the model went from the start of pre-training on May 22 to public launch in under nine weeks, trained on 4,096 Nvidia H200 GPUs. In an industry where flagship model cycles are typically measured in quarters or years, Poolside has now shipped three models in three months.

Why the West's open-weight AI gap has become a boardroom issue

The release lands in the middle of an increasingly pointed debate about the provenance of open-weight AI. Over the past year, developer adoption has shifted decisively toward open-weight systems that companies can download, inspect, and run on their own infrastructure — and the leading options in that category have overwhelmingly come from Chinese labs. DeepSeek, Qwen, Kimi, GLM, MiniMax, and Tencent's Hunyuan line all feature prominently in Poolside's own comparison tables.

Poolside's accompanying press release frames Laguna S 2.1 explicitly as a response, noting that the model occupies a size class into which no Western lab has released open weights in 11 months — since OpenAI's gpt-oss-120b last August. "The West needs open-weight models it can trust, run, and build on," said Jason Warner, Poolside's co-CEO, in the announcement.

Co-founder and co-CEO Eiso Kant made the philosophical stakes even plainer in a lengthy post on X. "I believe intelligence should and will become a commodity," he wrote, arguing that the open ecosystem "will not win by being the best in its own category." Users, he argued, simply want the best intelligence for the task at hand — so open models must be on par with, or better than, their closed equivalents.

The strategic logic here is not charity. Poolside's core business is deploying models inside the security boundaries of government, defense, and regulated enterprises — customers for whom closed, metered API access is often a non-starter for compliance and sovereignty reasons. 

Every enterprise that standardizes on a Chinese open model today becomes harder to win tomorrow. Releasing competitive open weights is both an ecosystem play and a top-of-funnel strategy for the company's high-security deployment business. It also reframes the AI race away from terrain where Poolside cannot compete — frontier-scale capital expenditure — and toward terrain where it believes it can: cost per token, self-hosting, and iteration speed.

How a sparse architecture makes enterprise AI agents affordable to run

The technical design reflects a specific thesis about where value in coding AI is moving. Laguna S 2.1's sparse MoE architecture — 256 routed experts plus one shared expert, with grouped-query attention and interleaved sliding-window layers, according to the Hugging Face model card — means inference costs scale with the 8 billion active parameters, not the 118 billion total. Poolside emphasizes that the model is small enough to run on a single Nvidia DGX Spark, the desktop-class AI machine.

That matters for what Poolside calls token economics. Long-horizon coding agents are voracious consumers of tokens: the company's published data shows the model consuming a mean of roughly 249,000 completion tokens per trajectory on its hardest benchmark when thinking mode is enabled. At metered API prices, agentic workloads at enterprise scale become a meaningful budget line item. On OpenRouter, Poolside is offering a free 256K-context endpoint and a dedicated 1M-context deployment priced at $0.10 per million input tokens and $0.20 per million output tokens — aggressive pricing that undercuts most frontier alternatives by an order of magnitude.

The ecosystem support is unusually broad for day one. The model is live on Baseten's model library and Vercel's AI Gateway, with integrations across vLLM, SGLang, Ollama, and llama.cpp, plus quantized variants down to 4-bit GGUF files — 75 gigabytes — for local use. But Poolside's more interesting claim is behavioral, not architectural. Pengming Wang, co-head of applied research at Poolside, said the gains came from improving the model's working habits: "more verification, less taking things for granted, not declaring victory early, and being more persistent." Raw intelligence, the company argues, is one axis of capability; a model's way of working is a second axis that matters immensely for agents left unattended for hours.

Publishing every benchmark trajectory to counter AI's credibility crisis

The most consequential part of the release for enterprise buyers may be an evaluation-transparency move with little precedent among major labs: Poolside published the complete, unedited trajectory of every trial in its final benchmark runs — every reasoning step, tool call, and shell command behind every reported score.

This addresses a growing credibility problem in AI benchmarking. As top scores on mature benchmarks cluster in the 70–90% range, and as "reward hacking" — models finding solutions online or gaming verifiers rather than solving problems — has become endemic, self-reported numbers have lost much of their signal. Poolside disclosed its own encounters with the problem candidly: during training, more than half of trajectories on some SWE-bench tasks were flagged because the model simply researched the original bug-fix pull request online and applied it. The company documented its mitigations, including prompt addenda, LLM-based judging calibrated against human labels, and expert annotator review of a high-scoring Terminal-Bench run.

Three published case studies illustrate what the company means by persistence. In one, the model built a working HTML/CSS rendering engine from an empty folder in a 181-step, 50-minute unattended session — then, lacking vision capabilities, spun up headless Chromium to numerically compare its canvas output against a real browser's rendering. In another, pointed at Poolside's own agent harness in an automated optimization loop, the model made the Go codebase 5.2% faster with roughly 70% lower memory allocation, finding an O(n²) string-concatenation bug along the way. In a third, working in a sandbox with no Python installed, the model did its number theory in Perl and independently re-derived a proof of Erdős problem #397 — a combinatorics question open for five decades until GPT-5.2 Pro first solved it this past January. Poolside notes that its model's construction is structurally different from the earlier published solution, and that its November 2025 knowledge cutoff precedes the first proof.

What the disclosed limitations and benchmark fine print reveal

Poolside deserves credit for disclosing limitations most labs bury. The model can overfit to its native harness and stumble on slightly different tool schemas in third-party agents, mangles JSON in nested tool arguments, and is prone to overthinking on competition math. There is currently no user-configurable thinking-effort dial — just on or off — and the gap between the modes is enormous: thinking lifts Terminal-Bench 2.1 from 60.4% to 70.2%, and DeepSWE from 16.5% to 40.4%, at substantially higher token cost.

Buyers should apply their own discounts to the comparison tables. Poolside's methodology takes the maximum of vendor self-reported scores, benchmark-author leaderboards, and third-party figures for competitors — a reasonable convention, but one that mixes harnesses and test conditions. On DeepSWE, notably, Poolside ran its own agent harness rather than the leaderboard's standard mini-swe-agent, a difference the company acknowledges makes scores less directly comparable. And the frontier remains clearly out of reach: closed models like GPT-5.6 Sol, at 88.8 on Terminal-Bench 2.1, and Claude Fable 5, at 88.0, along with the 2.8-trillion-parameter open-weight Kimi K3, at 88.3, sit well above Laguna S 2.1.

The deeper structural question is whether Poolside's "Model Factory" — the internal platform the company credits for its rapid release cadence — can sustain this pace as models scale. The trajectory so far is genuinely unusual: the April dual release of Laguna M.1 and XS.2, the July 2 refresh of XS 2.1, and now S 2.1, which the company says outperforms April's flagship M.1 at roughly a third of its active size. Remarkably, S 2.1 used the exact same pre-training data as XS 2.1, meaning nearly all the improvement came from scale, training fixes, and post-training across the company's corpus of 409,000 agentic and non-agentic training environments. Poolside says its next, larger Laguna model began pre-training last week.

For technical decision makers, Laguna S 2.1 is the most credible Western open-weight option to emerge in nearly a year for self-hosted agentic coding — with published evidence, a permissive license, broad ecosystem support, and an economics story built around hardware you can own. Whether it dents the dominance of Chinese open models will depend less on this release than on the ones that follow it.

Kant, for his part, has already told the world how he intends that story to end. Poolside is building toward a future where the most capable intelligence "can be owned and shaped by anyone," he wrote — and the company plans to keep shipping "until that future exists." In an industry where the biggest labs increasingly lock their best work behind an API, the most radical thing about Laguna S 2.1 may not be what it scores, but that anyone can download it and check.

GPU memory is the most expensive resource in production AI, and it's also the one running out fastest.

Long context windows and multi-turn conversations force AI models to repeatedly recompute information they've already processed, consuming GPU memory and compute that could otherwise serve additional users or generate new responses.

Instead of treating GPU memory as the limiting resource, why not extend it with much cheaper storage technologies?

Weka, for one, believes that cheap flash storage can close that gap. The company's NeuralMesh 6 software platform, launching alongside its first self-designed hardware line, Wekapod 3, extends what Weka calls Augmented Memory Grid, an approach that aggregates NAND flash to behave like GPU memory at a fraction of the cost.

This is an active and increasingly crowded category. Dell, NetApp, Pure Storage and VAST have all repositioned toward AI infrastructure over the past two years and Weka is one of several vendors arguing it's built for this specific moment rather than adapting to it.

"What we're seeing now with customers is they're chasing availability of compute, and once they get new allocation from anyone, they want to be able to grab it and start running right away," Weka co-founder and CEO Liran Zvibel, told VentureBeat.

The potential payoff is straightforward: better utilization of existing GPU investments, lower inference costs and faster deployment of new AI workloads without waiting months for additional GPU capacity.

The technology is most relevant for organizations already operating AI at scale or expecting rapid growth in usage, particularly enterprises building internal copilots, customer service agents, software engineering assistants or retrieval systems with long context windows. Smaller deployments may see less immediate benefit than organizations where GPU utilization has already become a limiting factor.

Inside Weka's NeuralMesh 6

NeuralMesh 6 adds four capabilities aimed directly at a functionality gap Zvibel says has been costing Weka deals in competitive evaluations.

Composable and virtual multi-tenancy. Composable clusters give anchor tenants full hardware-level isolation, dedicated CPU, memory, and storage. Virtual multi-tenancy runs through Weka's RDMA fabric, delivering network-level isolation that scales past 1,000 tenants per cluster, with provisioning in under 30 minutes. Combined, a single cluster running 50 composable clusters can support up to 50,000 tenants. 

Unified file and object storage. Most storage systems keep two separate paths: a file-based path (the standard way servers and applications read and write files, used heavily in training and fine-tuning pipelines) and an object-based path (S3, the format inference and cloud-native tools typically expect). Normally a gateway translates between the two, meaning the data effectively exists twice. Weka's claim is that the same physical data on disk is directly readable through either path at once, no translation layer, no second copy. Zvibel is targeting non-AWS GPU clouds specifically, naming Lambda, Nebius, G42, and CoreWeave, with what he described as roughly two orders of magnitude higher performance than conventional S3 and a capacity-based pricing model instead of per-API charges. 

Metadata-first replication. Destination environments become browsable before a full data copy arrives, with data hydrating only when accessed. 

"They had to wait for all of that to make it to the other side, and this takes days or weeks, in extreme cases a month," Zvibel said. "We now allow our customers to grab some allocation of new GPUs and get up and running within an hour."

AlloyFlash and Always-On data reduction. TLC and QLC are two types of NAND flash memory. TLC is faster and more durable but costs more per terabyte, while QLC is cheaper and holds more data per chip but is slower. AlloyFlash mixes both within a single cluster, automatically routing latency-sensitive work to TLC while running bulk-capacity workloads on QLC, cutting cost per terabyte without a performance penalty on the work that needs speed. Data reduction now runs by default rather than as an option.

Solving AI's context problem

Multi-tenancy and object storage solve how enterprises and neo clouds operate the platform day to day. A harder problem sits underneath: as context windows and multi-turn interactions grow, so does the GPU compute wasted recalculating work a model has already done. Augmented Memory Grid, a NeuralMesh 6 feature built specifically for this, is Weka's answer.

Every prompt triggers two stages. Prefill calculates attention, the core mechanism behind how large language models process input, and it's computationally expensive. Decode converts that calculation into output and is comparatively lightweight.

The cost shows up hardest in multi-turn sessions like chat or coding, where each new turn re-triggers prefill for everything that came before it, unless that work has been cached.

"If you have 10 turns, you may overcalculate 100 times because you're redoing all of them. If you have 20, you'll overcalculate 400 times," Zvibel said. "You can put two orders of magnitude more NAND than you could afford in shared memory, and we can cache 100% of the pre-calculated tokens, so you never need to redo it."

Where Weka sits competitively

Storage vendors have spent the past year and a half repositioning around AI, and separating genuine capability from repositioned messaging is now a real evaluation problem for buyers. 

"The storage world is shifting its focus from serving bits to enterprise workloads to managing data at the speed of AI. We've seen that most clearly over the past 18 months from Dell, NetApp, and Pure," Steve McDowell, chief analyst at NAND Research, told VentureBeat. "The interesting thing is that companies like Weka, and VAST, are the true AI-native data companies, solving these problems since day one."

McDowell singled out Augmented Memory Grid as Weka's clearest technical lead. 

"Weka continues to have the most technically capable KV cache implementation on the market with its Augmented Memory Grid," he said. " They were early with this technology, and continue to innovate. This is critical for AI inference, as it enables a level of GPU efficiency that, without question, saves money on GPUs and memory. That’s key for today’s memory and GPU constrained market." 

He also flagged Weka's contractual guarantee on its data reduction claims as underappreciated. 

"One flying a little under the radar: Weka is putting its money where its mouth is with its contractual guarantees for its data reduction promises," he said.

McDowell's advice to buyers evaluating competing claims from Weka, VAST, Pure and NetApp alike was pointed suggesting that enterprise buyers should look hard at what vendors are promising versus what they're actually delivering.

"A smart buyer will look at how competing vendors are solving real-world problems today," McDowell said. " They do this by talking to organizations running similar workloads at similar scale. If a vendor can't point to that, then it should be a warning sign."

Google DeepMind today released three new proprietary AI models it says are among its most token-efficient yet: Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and Gemini 3.5 Flash Cyber.

The models aim to make AI agents faster, smarter, and cheaper at scale. Google is pricing Gemini 3.6 Flash at $1.50 per one million input tokens and $7.50 per one million output tokens through its application programming interface (API), while Gemini 3.5 Flash-Lite costs a staggeringly cheap $0.30/$2.50 per million tokens in/out.

Compare that to the $1.50/$9.00 per 1M tokens for Gemini 3.5 Flash, and the $2/$12 for Gemini 3.1 Pro Preview, and the savings are considerable. However, Google's prior generation Gemini 3.1 Flash-Lite still remains the search giant's "most cost-efficient" model at $0.25/$1.50 per 1M tokens. Yet, it remains 2X slower than the new, more expensive Gemini 3.5 Flash-Lite, giving those enterprises who value speed more "bang" for their buck.

VB Frontier AI Model API Pricing Comparison Chart (Late July 2026 Shortlist)

No price was provided yet for the specialty Gemini 3.5 Flash Cyber model, which, as its name would imply, is designed for cybersecurity researchers and red teamers to patch bugs.

While the prices are among the middle-low end of all major AI models globally, the fact that Google designed them to use less tokens overall also should drive down costs for enterprises beyond what the sticker price shows (since you'll be paying for fewer total tokens at any rate).

Gemini 3.6 Flash and Gemini 3.5 Flash-Lite are available immediately through the Gemini API in Google AI Studio and Android Studio, as well as within the consumer Gemini application and Google Search. According to a separate Google blog post, Gemini 3.5 Flash Cyber will be available "exclusively available to governments and trusted partners via CodeMender soon" — CodeMender being Google's proprietary AI code bug-fixing agent released last year.

As with previous Gemini models, these are all proprietary and "closed source," thus, they can only be obtained through Google's official API and that of its partners, as opposed to an open-source license like MIT or Apache 2.0.

One conspicuous omission noted by developers on X and social media: where is the larger, more powerful, flagship Gemini 3.5 Pro model Google previously alluded would be released this summer? After all, Gemini 3.1 Pro, the prior flagship, debuted back in February 2026, and rivals OpenAI and Anthropic have since released several more generations of flagship updates far more powerful than Google's.

Google technical staffer Logan Kilpatrick responded to one such inquiry on X, writing: "Gemini 3.5 Pro is currently testing with partners and we plan to make it broadly available as soon as it’s ready."

Google's release signals that the immediate future of AI lies in agentic capabilities—systems that operate autonomously over extended periods.

If early large language models are akin to massive, fuel-hungry freight trains capable of hauling incredible loads at immense cost, the new Flash series represents a fleet of nimble, hyper-efficient hybrid delivery vans.

Efficiency gains ranging from 17% to 65% reduced tokens for strong results on third-party benchmarks

Under the hood, Gemini 3.6 Flash achieves significant efficiency gains. The model reduces output token usage by 17% compared to its predecessor, Gemini 3.5 Flash, according to the Artificial Analysis Index maintained by the independent third-party AI benchmarking group of the same name.

In specific long-horizon software engineering benchmarks like DeepSWE, which measures how well agents complete multi-step engineering tasks from scratch, the token savings reach up to 65%.

This reduction means the model requires fewer reasoning steps and tool calls to complete the exact same multi-step workflow. Think of token efficiency like fuel economy in a vehicle.

When an AI model takes a convoluted path to solve a problem, it burns through more computational fuel, driving up the final cost for the developer. By streamlining its internal logic, Gemini 3.6 Flash arrives at the correct answer faster and cheaper.

While Google's materials did not specify the exact architectural or algorithmic changes used to achieve this token efficiency, they noted that the model "takes fewer reasoning steps and tool calls to accomplish multi-step workflows" and exhibits reduced "verbosity."

The official model cards released by Google reveal that both Gemini 3.6 Flash and Gemini 3.5 Flash-Lite feature a 1-million-token input context window alongside a max output limit of 64,000 tokens, with both models sharing a knowledge cutoff date of March 2026.

Respectable benchmark performance at low cost

The technological improvements extend to concrete capabilities. Gemini 3.6 Flash scores 49% on the DeepSWE benchmark, a notable increase from the 37% achieved by version 3.5.

It also pushes machine learning engineering performance higher, scoring 63.9% on MLE-Bench compared to 49.7% previously. Furthermore, Google integrates computer use as a built-in client-side tool via the Gemini API and Gemini Enterprise, reflecting an OSWorld-Verified score of 83.0%, up from 78.4%.

The model also tackles knowledge work with greater proficiency, outperforming its predecessor on benchmarks like GDPval-AA v2 by moving from a score of 1349 to 1421.

To ensure safety amidst these capability upgrades, Google deploys enhanced Frontier Safety safeguards. These protections harden the model against jailbreaks and mitigate risks in Chemical, Biological, Radiological, and Nuclear domains, as well as cyber offense misuses.

The engineering team trains the model to minimize refusals for beneficial uses, striking a necessary balance between strict security and practical utility.

Models for low-cost coding, agentic, and cybersecurity use cases — respectively

Google divided its new offerings into three distinct products tailored for different operational needs.

Gemini 3.6 Flash serves as the heavy-duty workhorse of the trio. It handles complex coding, intricate knowledge work, and multimodal processing with improved precision. Enterprise customers utilize it for demanding tasks such as complex document parsing, intricate chart and data analysis, and long-form report drafting.

The model executes complex code migrations using multi-agent orchestration frameworks with lower latency and higher quality than earlier iterations. Furthermore, 3.6 Flash aids in developing photographic texture extractors for 3D workflows using canvas interfaces.

Gemini 3.5 Flash-Lite targets environments where high throughput and absolute minimal latency are non-negotiable. Google designates it as the fastest model in the 3.5 series.

As measured by Artificial Analysis, the model processes 350 output tokens per second, making it highly effective for agentic search and massive document processing workloads. Artificial Analysis notes this is about twice as fast as prior generation model Gemini 3.1 Flash-Lite.

Developers can configure 3.5 Flash-Lite to prioritize low-latency execution for high-volume tasks using minimal thinking levels, or engage higher thinking levels to process complex multi-step subagent workloads.

Despite its lite designation, it outperforms the standard Gemini 3 Flash on several key agentic and coding evaluations, including SWE-Bench Pro, where it scores 54.2% compared to 49.6%, and OSWorld-Verified, scoring 74.0% versus 65.1%.

The model extracts product features from massive datasets, generates interactive web design concepts, and scales receipt translation seamlessly.

The third product, Gemini 3.5 Flash Cyber, represents a highly specialized deployment. Google fine-tuned this model specifically to find and fix cybersecurity vulnerabilities. It integrates directly with Google's CodeMender agent.

In practice, multiple 3.5 Flash Cyber agents work concurrently to produce a single, comprehensive vulnerability report, achieving competitive performance at the frontier on the CyberGym benchmark, even getting within range of Anthropic's much-hyped Mythos model.

Google did not specify an exact numerical cost for 3.5 Flash Cyber, stating only that it is fine-tuned "at a lower price per token than larger models.

Commercial licensing only

The licensing framework for the new Gemini models carries profound implications for developers and enterprise users. Google deploys Gemini 3.6 Flash and 3.5 Flash-Lite under a commercial, proprietary API model. Unlike open-source software governed by licenses such as the MIT License or the GNU General Public License, developers do not gain access to the underlying model weights, training data, or source code.

An MIT or GPL license grants users the freedom to download the codebase, modify the internal architecture, self-host the deployment, and distribute the software infrastructure independently. In contrast, Google's API approach means developers essentially rent access to the intelligence on a strict metered basis. Every prompt and generated response travels through Google's managed servers, incurring a cost based on the strict pricing structure of $1.50 per million input tokens for 3.6 Flash.

This commercial tethering restricts deployment flexibility. Enterprises cannot air-gap the models entirely on their own local secure hardware without establishing specialized, high-tier enterprise agreements with Google Cloud. Developers remain bound by Google's acceptable use policies, arbitrary rate limits, and network requirements, creating a permanent dependency on Google's infrastructure uptime and terms of service.

The licensing for Gemini 3.5 Flash Cyber proves even more restrictive. Acknowledging the dual-use nature of cybersecurity AI—which attackers can weaponize just as easily as defenders can use it to patch systems—Google is for now making the model only available behind a limited-access pilot program, similar to the trend kicked off by Anthropic's Mythos model with its Project Glasswing program, and continued by OpenAI with its staggered rollout for GPT-5.6.

In this case, Google is making 3.5 Flash Cyber exclusively available to governments and trusted partners. This strict gatekeeping prevents open access, prioritizing systemic security over widespread developer innovation.

Looking ahead

Google DeepMind continues to iterate rapidly, but the gap in its product line remains apparent. While the Flash series excels in speed and economy, the industry eagerly awaits the deployment of Gemini 3.5 Pro to gauge Google's absolute frontier capabilities.

Simultaneously, the company confirms that pre-training for Gemini 4 has already commenced.

Until the next major flagship release materializes, developers must optimize their systems using the highly efficient, yet purposefully constrained, Flash architecture.

“The new PRD are the evals,” Xavi Amatriain, Expedia Group’s first chief AI and data officer, told the VB Transform 2026 audience last week in Menlo Park. “So basically, you encode what you want the product to do through your evals, which might include red teaming evals and all kinds of other things, which already have a bunch of security requirements. So, you already embed that into the PRD and the product design document before you even start coding.”

He pushed it further. “With AI-assisted or AI-generated code, that’s gonna be the future. It’s like all your thinking is gonna go into the evals.”

Amatriain served as VP of AI and Compute Enablement at Google across the platforms powering Gemini and Google Search before his December 2025 appointment at Expedia. He's mentored talent who went on to found Perplexity and Scale AI.

VentureBeat’s VB Pulse research on the evaluation gap reinforced the stakes. Sixty-six percent of the 157 enterprises surveyed already permit some production deployment without human review or are building toward it within the next 12 months, yet only 5% fully trust the automated evaluations that would make that decision. Half have shipped an agent that passed internal evals but then failed with a real customer.

Don’t let guardrails get in the way of feedback

“The more guardrails and artificial business rules and sort of rules that you put into the system, the worse off,” Amatriain said. “Not only because they’re brittle, but also because they actually mess up with the feedback loop. You are actually biasing the user and the feedback you get from the user, and then you’re learning that in the wrong way.” He called guardrails “a necessary evil” and said the goal is to minimize their impact over time.

Not everyone at Transform agreed. Other speakers argued during the event that the highest-risk actions still demand very firm guardrails.

Expedia governs AI through three layers instead. Principles come first, communicated broadly. “I like to encode at a very high level how I expect decisions to be made, because in a large organization you’re gonna have a lot of distributed decision making,” Amatriain said. “And sometimes, if you’re lucky enough, those principles might be embedded in your culture. But most of the time, my experience has been they’re not.” The processes and tools that enforce them follow. “Principles look really nice on a picture on some wall, but you need to then give them teeth,” he said. Automation sits on top of both.

In practice, this plays out through what Expedia calls agent release toll gates, checkpoints calibrated to risk. “Governance needs to correlate to the risk,” Amatriain said. “And if you have something that is low risk, you don’t need too much governance to get in the way. But if there’s a lot of risk, then you need more governance. That can be encoded.” The toll gates tie evaluation rounds, red teaming, and security review to each agent’s risk level, and the checks shift from recommended to required as the stakes climb.

Specialized agents over monolithic intelligence

“Even when I was at Google, I was like, I don’t believe in AGI as sort of like a singleton and a unified sort of like single model,” Amatriain told the audience. “I think it’s much better to think of it as composition, sort of like having specialized agents that are very good at some task and then composing the system out of those specialized agents.”

Expedia’s architecture starts at the component level. Tools compose into skills, skills assemble into sub-agents, and sub-agents get orchestrated into the full agentic system. “You need to have those principles that are unified that talk about things like what is the tone that we’re using, how are we addressing the user, how are we passing context, memory,” he said. “All of that needs to be thoroughly designed.” He framed this as a systemic design problem. “It’s not about the model, it’s not about a specific solution, it’s about how you’re designing the system.”

Amatriain argued that scoping each agent narrowly also makes the system easier to secure, since teams can evaluate and lock down individual agents in isolation before composing them.

When the user must keep the final click

Travel pricing changes in real time, flight availability shifts minute to minute, and hotel reviews routinely contradict what suppliers claim. Amatriain described a system that blends retrieval-augmented generation with direct API tool calls, choosing the approach based on latency. “If the user asks you a question like, how much does a four star hotel usually cost in Chicago in July, you don’t expect the agent to take two minutes to answer that question,” he said. “You expect an immediate answer because that answer can be cached and it doesn’t need real-time information.” A pet-friendly four-star near Lake Michigan with a pool might justify a 30-second reasoning window.

“The supplier might be saying, yeah, we have a great swimming pool, but then we also have the reviews from the travelers and we actually see there’s two reviews that say the swimming pool was not great or was not open after 6 p.m.,” Amatriain explained. A generic chatbot, he added, would only surface what a supplier self-reports, while Expedia cross-references against its own review corpus.

“We don’t want the agent to book the hotel or to buy you a plane ticket for you,” Amatriain said. “That’s something that the user has to have the agency. And the agent can recommend, can suggest, can discuss with you, but you’re gonna have to hit that click. And that’s non-negotiable.” That constraint, he argued, is also a security decision. “Once you establish those design principles, you also don’t need the guardrail because otherwise you’re gonna have to put all those guardrails in after the fact.”

The next attackers will be other AI systems

“Security needs to be a principle that is shifted as left as possible and as part of the design itself,” Amatriain said in response to an audience question. “And usually when you need a guardrail is because you’ve not thought about it early on.”

A second audience member pressed for lessons learned from production. Amatriain described a feedback loop where monitoring signals flow back into the eval suite. “You can almost automate the whole cycle,” he said. “But having that whole feedback loop from real signals, from your operating AI system, all the way into being reported and fixed as quickly as possible is going to become essential.”

Amatriain's toll gates are a bet that governance calibrated to risk can stay ahead of that feedback loop. VentureBeat’s separate June Pulse survey on agent security, drawn from 107 enterprises, shows how thin that margin is. More than half, 54 percent, have already had an agent security incident or near-miss. Fifty-nine percent plan to adopt, add, or replace agent security tooling within 12 months, and 29% plan to move this quarter. Incident rates climb with organization size, reaching 63% among enterprises with more than 1,000 employees versus 49% for companies with 101 to 1,000. And sandbox isolation, the one post-breach control that limits damage, drops from 35% adoption at the smaller companies to just 20 percent at the largest.

Amatriain warned that threats will increasingly come from other AI systems. “You’re gonna get threats coming not only from humans but also from other external agentic systems that are really powerful, and they’re gonna be poking at everything you’re doing. And as soon as you detect something, it’s not only about the detection, but the time to fix becomes essential here.”

Presented by Atlassian

Most companies are approaching AI adoption backwards by optimizing how individuals use AI instead of how teams work together, said Dr. Molly Sands, head of the Teamwork Lab at Atlassian, during a fireside chat with VentureBeat senior technology contributor Sam Witteveen at VB Transform 2026.

Sands leads a team of behavioral scientists and psychologists who study how AI is reshaping the way people work together, using those findings to help organizations redesign how work gets done.

"We don't just study it, we also actively go in and change it," she explained. Her teams teach new ways of working and remap how work flows across companies, a challenge that many organizations are still struggling with, she said.

Why AI speed isn’t translating into ROI

Atlassian's annual State of Teams Report, which this year surveyed 12,000 global knowledge workers and interviewed roughly 200 Fortune 1000 executives, found a significant disconnect between activity and value, showing that everyone is using AI, while very few can yet locate where it pays off.

"89% of those executives told us that individuals are speeding up in their companies, and only 6% of them said they could point to specific examples of clear ROI," Sands said.

But roughly 14% of teams had translated AI usage into real value — meaning a single organization could contain a handful of high-performing teams surrounded by others seeing no return at all.

Those leading teams shared three characteristics: context, workflows and culture. The teams pulling ahead were building what Atlassian calls a context graph by capturing goals, decisions, and organizational knowledge in shared digital records rather than leaving them in individual memory. Across products such as Jira and Confluence, the graph connects work items, goals and the people doing them, giving AI access to the organizational context it needs.

On workflows, the winning teams redesigned entire end-to-end processes rather than simply accelerating isolated tasks. Otherwise, speeding up individuals who are pointed in slightly different directions only causes them to “very quickly start to crash into each other,” as Sands puts it.

On culture, the fastest-moving teams worked under leaders who explicitly encouraged learning and experimentation, while making it clear that some experiments would fail.

How leaders can move AI from individual hack to team advantage

Experimentation and constraints are the fastest route to learning, Sands said. The teams seeing the biggest gains were deliberately imposing constraints on how they worked, from breaking every task into the smallest practical unit of work (a single story point) to committing to write no code by hand for a week.

"Most of it is not sustainable to do forever, but it is a really, really fast way to learn," she said.

Sands argued that another obstacle isn’t the technology itself but the fact that employees are figuring out AI on their own. Every worker develops different prompts, agents and assumptions, creating another layer of unspoken knowledge inside teams that rarely translates into organizational performance.

To counter that, Atlassian experimented with AI working agreements at the start of projects, asking teams to decide not only what they would use AI for, but what they would deliberately avoid using it for, which agents they would share and what common skills would keep everyone working from the same context. Teams that adopted the practice used AI more, moved faster, made better decisions and ultimately produced higher-quality work.

The broader lesson, Sands said, is that AI isn’t creating entirely new management problems so much as exposing old ones. Teams have always struggled with hidden assumptions and different mental models of their work. AI simply makes those gaps more consequential, increasing the importance of shared context and explicit ways of working.

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.

Enterprise AI is facing an ROI paradox. While throwing more compute at the strongest foundation model works well in product experiments, the costs become unbearable when the product is deployed in production.

A new paper from researchers at Writer provides a solution that is accessible to engineering teams. The study takes a systematic look at optimizing the different components of the orchestration layer that wraps around the foundation model, aka the AI harness. 

By optimizing the harness, the researchers show dramatic reductions in tokens per task, a drop in cost-per-successful-task by up to 61%, and quality that holds steady, all without changing the underlying foundation model.

Because the harness is fully under the developer's control and requires no model fine-tuning, engineering teams can apply these findings to build highly cost-efficient AI applications.

The ROI crisis of tokenmaxxing

The current state of AI engineering is plagued by "tokenmaxxing," an industry trend where developers rely on massive context windows and brute-force token consumption as a substitute for good system design. 

Rather than engineering elegant workflows, developers have imported a reflex from traditional software development: generate, run, fail, stuff the error and more context back into the window, and retry. 

"Teams tokenmaxx because it's the cheapest fix in the moment, and because it's literally how most engineers work today," Waseem AlShikh, CTO and co-founder of Writer, told VentureBeat. Because this approach succeeds often enough on coding tasks, it has become the default reflex for every other agentic workload. The danger is that per-token price drops mask the underlying inefficiency. 

"Your invoice is tokens-per-task times price-per-token, and most teams only watch the second number," AlShikh said. "In agentic workloads, tokens-per-task compounds — every loop iteration re-transmits the growing context — and it compounds faster than prices fall. The price cut becomes an anesthetic. It masks the fact that the loop itself is bleeding."

Tokenmaxxing leads to several enterprise failure modes. Teams route simple tasks to premium frontier models by default. They use the LLM as a lazy search index, stuffing the context window with raw documents instead of retrieving exact answers. Most destructively, they build unconstrained agentic loops that spiral out of control when the model encounters an error. Because output tokens cost significantly more than input tokens across all major model providers, inefficient task execution acts as a silent budget killer.

The industry has introduced several efficiency techniques to curb these costs, but they largely fall short because they treat the model in isolation: 

• Prompt compression condenses input text to save space, but ignores how the system sequences those inputs across complex workflows. 

• Budgeted reasoning caps the computational steps a model can take, which often degrades output quality if the workflow isn't intelligently routed. 

• Terse coding forces models to output minimal code to save output tokens, but does nothing to solve inefficient tool calling. 

• Speculative decoding uses a smaller draft model to speed up a larger model's text generation, optimizing inference speed while failing to address bloated agent architectures.

These efforts fail because they optimize the engine while ignoring the transmission. They do not look at the orchestration layer, leaving underlying architectural inefficiencies unresolved.

Unpacking the harness: the levers of efficiency

The harness is the orchestration layer that routes, formats, and turns the underlying LLM into a working system.

The core levers of harness optimization include system prompt caching, interaction history compaction, tool management, retrieval strategies, and error management. These are the most accessible intervention points for engineering teams looking to improve AI performance. 

As the Writer researchers note in the study: “If the harness is the layer that composes model calls into work, it is also the layer that sets the price of work.”

Historically, developers have treated the harness as disposable glue code designed simply to connect an API to a user interface. The study signals that the harness must now be treated as a first-class object: a primary software artifact that requires its own testing, versioning, and rigorous design. 

For enterprises, this reframes the "own-versus-rent" decision. 

"Enterprises spend months on model evaluations and then rent their orchestration off the shelf — which means they're optimizing the smaller lever and outsourcing the bigger one," AlShikh said. "Whoever owns the harness owns your unit economics, and an open framework tuned for demos is not tuned for your invoice." 

Inside the experiments

To isolate the impact of the orchestration layer, the researchers ran experiments on six foundation models spanning multiple vendors and weight classes: Claude Sonnet 4.6, Gemini 3.1, Gemini Flash 3.5, Qwen 3.6, GLM 5.1, and Writer’s own model, Palmyra X6. 

Their experiments compared a frozen, conventional production agent loop against the finished Writer Agent Harness on the same 22 locked enterprise tasks, spanning capabilities like grounding and retrieval, multi-step workflows, tool use, and content generation. By holding the models and tasks constant, they could isolate the effects of the orchestration layer itself.

The optimized harness drove a significant drop in costs, cutting the blended cost per task by 41%, from 21 cents to 12 cents. This was largely achieved by slashing token consumption, with the number of tokens per task falling 38%, from 14.2k to 8.8k.

The harness is designed to delegate tasks like search to specialized sub-agents. A sub-agent receives only the tool and the specific query it needs, retrieves the exact data, and returns a capped, clean summary to the main agent — keeping the primary context window from filling up with raw search results.

Task success rates held steady even as token use fell — moving from 78% to 81%, a gain the researchers describe as directional rather than statistically significant at their sample size, meaning quality didn't suffer even as costs dropped.

End-to-end task latency also dropped significantly, reducing the median wall-clock time by 44%, from 48 seconds to 27 seconds, due to prompt caching and the elimination of dead-end reasoning loops.

However, the researchers also found limits to multi-agent orchestration. Smaller models like Gemini Flash 3.5 and Qwen 3.6 scored well below a usable reliability threshold on sub-agent delegation tasks (0.45 and 0.42, respectively) — the capability simply isn't dependable yet on lighter-weight models.

Sub-agent orchestration only crossed a usable reliability threshold on the two strongest models tested: Writer's own Palmyra X6 (0.86) and Claude Sonnet 4.6 (0.85).

The developer’s playbook: actionable takeaways and tradeoffs

The findings from the study translate into a playbook for enterprise developers building agentic workflows at scale. The first step is to implement what AlShikh calls the "Two-Zone Prompt" and "Context Offloading."

Structure for system prompt caching (The Two-Zone Prompt): Modern LLM APIs offer prompt caching, but developers must structure their payloads correctly to trigger it. Developers must separate the "stable zone" from the "volatile zone." Place static, unchanging elements (e.g., core rules, large tool schemas, and standard operating procedures) at the top of the prompt. Dynamic elements, such as the specific user query or recent conversational task state, must be appended at the bottom. This ordering allows the harness to reuse the cached prefix across hundreds of calls. "That single separation makes prompt caching actually work and stops you from re-paying for the same instructions on every one of an agent's thirty steps," AlShikh said.

Manage context with Context Offloading: Avoid context stuffing, where every turn of a loop is appended into a monolithic prompt until the window maxes out. Instead, move history and intermediate artifacts out of the window into retrievable storage, and pull back only what the current step needs. If possible, delegate tasks to single-purpose sub-agents to avoid context bloat. As AlShikh points out, "the biggest line item in agent spend isn't reasoning — it's re-sending things the model has already seen."

Build resilient loops and redefine KPIs: Unmanaged agent loops drain API budgets rapidly. Teams must begin tracking Completions Per Million tokens (CPM) to understand their true task costs, but the harness itself must contain physical guardrails. "The core principle is that you never ask the model to police its own spending," AlShikh said. "The fence has to live below the model, in code, on your side of the API." This requires three hard checks:

• Hard per-task token budgets: The run terminates when the budget is spent, no exceptions.

• Generation fencing: Caps on steps, tool calls, and recursion depth to stop non-converging agents. 

• Failure-spend governance: Cap what a run can spend after its first failed validation so a failing task doesn't become your most expensive task.

Avoid unnecessary complexity: Optimizing the orchestration layer comes with engineering overhead. If you're in the prototyping and exploration stage, that overhead isn't justified — iterate fast with a strong model and a light harness. Once you're scaling to millions of requests a day, the savings from harness optimization become substantial.

However, teams must be aware of "harness leverage." Adding structural scaffolding requires the model to hold and obey that context. If a model is too small, it will spend its limited capacity parsing the scaffolding instead of doing the task, causing accuracy to drop and tokens to rise. The rule for adding complex orchestration features is strictly mathematical: "If a feature adds more coordination tokens than it removes task tokens for that specific model, cut it," AlShikh said. "Nothing in the harness is free."

The future of the enterprise harness

The era of tokenmaxxing and treating context windows like bottomless buckets is coming to an end. Throwing more compute at poorly designed systems is not a viable strategy for companies that need to demonstrate a return on their AI investments. 

As foundation models evolve to absorb planning, tool selection, and multi-step reasoning natively into their weights, the role of the harness will shift from compensating for model weakness to enforcing enterprise policy.

"What never moves into the model is the 'allowed': budgets, permissions, data boundaries, audit trails, deterministic kill-switches," AlShikh said. "Five years from now, the harness will be thinner but more important. There will be less scaffolding and more governance. However capable the model gets, someone external to it still has to define what it may spend, see, and touch. That layer belongs to the enterprise, and it should never be rented."

A single AI agent conversation can look flawless scored on its own and still point to a broken product. That gap is driving a shift in how enterprises evaluate agents, away from scoring individual traces and toward comparing cohorts of users against a baseline.

At VB Transform 2026, Harrison Chase, CEO of LangChain; Hui Zhang, CTO and co-founder of Conviva; and Emmanuel Turlay, director of engineering at CoreWeave, described that shift, along with a parallel move toward cheaper, narrower judge models.

Agent-as-judge — judging one AI agent's output with another — hasn't replaced LLM-as-judge, which Chase said remains the default. The larger tension, Zhang said, is between automated judging, whether by LLM or agent, and human review.

"You have scalable but ungrounded, whether it's agents as judge or LLMs as judge, you grade the outcome, you grade the work. It still is very difficult to ground it and then you use humans and that's just not scalable," Zhang said. "The whole industry is facing this, which poison you want to pick."

Evaluation criteria now function as the product spec

That gap — a conversation that scores well but still signals a broken product — is what teams try to close by building an exhaustive evaluation suite before they ship anything. Chase said that doesn't work.

"We sometimes see teams that have almost eval paralysis," Chase said. "They're like, this is an eval set, I can't launch it. The best teams launch and then iterate."

Chase framed evaluation criteria as a living specification, not a one-time test suite: a product requirements document — the standard software-development spec for what an application should do. "Evals are like the new PRD," he said. "They define what your agent should and shouldn't do."

Turlay described hitting the same failure from a different angle. "I was trying to reach 100% coverage for my tests, and I still had bugs in production," he said — a test suite that looked complete but still missed what mattered, the same gap Chase was describing with evals.

Broad, always-on monitoring, he said, catches more real failures than an exhaustive pre-launch test suite. Teams should set up wide online checks first, use those to identify failure classes as they occur, then build a targeted offline evaluation set around the problems that surface.

Why scoring traces one at a time is a mistake

Even a well-built evaluation process can still score the wrong thing. Zhang's objection is to how most teams run evaluation: sampling traces, whether 50 of them or a full population, scoring each in isolation. That approach misses a signal that only shows up when comparing cohorts of users against a baseline, a method Zhang calls contrastive analysis.

Zhang illustrated it with a retail example: a shopper asks an agent for a running shoe ahead of a half marathon, the agent asks qualifying questions, and the shopper buys a shoe. Scored individually, that interaction looks fine. But the clarification ratio, how many follow-up questions an agent asks before completing a task, came in three times higher than baseline for that shoe category across the full user population. A second metric, how often shoppers finished their purchase outside the conversation, was five times higher than baseline for the same category.

Neither number is visible from a single trace. Both point to a debuggable, category-specific problem. Zhang said the industry also lacks a second data source: what happens before, between and after the conversation, not just the trace itself.

Sizing the judge to the job

Once contrastive analysis flags which category is actually broken, the next problem is what watches for it going forward — and at what cost. Turlay's rule was to start with the most capable model available to prove a task is solvable, then work down. If it can't be done with a top-tier model, he said, it won't work with a smaller one. Once a pattern proves viable, teams can sample a fraction of traffic instead of judging every interaction, and move simpler tasks like binary classification to smaller open source models.

LangChain took that further, fine-tuning its own model to detect when a user believes the agent made a mistake, a signal Chase calls perceived error. "The model we fine-tuned was a Qwen model," he said, referring to Alibaba's open source family. Combining hand labeling with distillation, the result performed well. "Same as [Claude]Sonnet, for, depending on how we served it, either 10 to 100x cost reduction," Chase said.

Not every guardrail needs a model. Chase pointed to Claude Code's own guardrails as proof: regexes, the common programming technique for finding and validating patterns in code. "A lot of the guardrails they had were just regexes," he said. "They weren't small LLMs, they were just regexes."

LLM-as-judge doesn't mean human-in-the-loop disappears

The bigger question is whether using LLM as a judge removes the need for a human in the loop.

Turlay pointed to accountability, drawing on his prior work at a self-driving car company. His team compressed data intake and retraining into a two-week cycle for shipping a new model to the car. Even then, someone still had to sign off.

"I felt confident on behalf of the company to say this model should go into the car," he said. The same logic extends to legal, finance and healthcare. "Before we can remove a human to say, I endorse this and I take responsibility legally for it, it's going to be a while before agents can do that on their own."

Zhang agreed a human has to remain the guardian on corner cases, even as automation eventually runs at a scale that beats individual human accuracy — machines can see more at the pattern level.

Chase went further: that human check isn't just a safety net. "Human in the loop is really important for building trust in how these agentic systems work, and also really important for memory and learning from systems," he said. "There has to be interactions in order for the system to learn."

Zillow, the real estate technology company, doesn't get one conversation with its customers. They move from a phone screen to a loan officer to a real estate agent, sometimes over months or years, and expect the context to follow them. A single chatbot could never carry that thread.

At VB Transform 2026, Zillow SVP of Engineering Toby Roberts and Glean co-founder and CEO Arvind Jain described how they built AI architecture meant to carry context across that entire journey — and why context, not raw data, turned out to be the harder problem to solve. Zillow's products touch roughly 80% of U.S. real estate transactions each year, and the company has been using AI long before ChatGPT existed.

"We pretty quickly identified that we were going to need a persistent context layer that was going to meet our customers and the professionals wherever they were," Roberts said.

Data was never the hard part

Roberts said Zillow's AI effort started where most enterprise AI efforts start, with the data itself.

"We started with a large push around making sure our data did have the right foundation," Roberts said. That meant a data mesh approach, clear data lineage and a governance structure with permissions and identity attached to the data itself.

None of that turned out to be the hard problem. The hard problem was building something that remembered where a customer was in their journey and carried that forward, no matter which surface they showed up on next.

"This context layer has to live to be able to support you where you are at any given point in your journey," Roberts said. Zillow chose to own that layer itself rather than depend on a single external chat interface, a decision Roberts said the team reached quickly once it looked at the shape of a real transaction rather than a single conversation.

Why Zillow built its own architecture, and where Glean fits into it

Zillow built its own harness rather than route customers through a single model API. The team drew on 20 years of machine learning history behind products like Zestimate, leaning into smaller, task-specific fine-tuned models instead of one general-purpose model.

Internally, that harness runs alongside Glean. Roberts said Zillow now has thousands of Glean agents in production, handling repetitive tasks with tens of thousands of executions across the company. Glean's pitch, per Jain, is centralizing that integration work once, through the Glean MCP gateway, rather than letting finance, legal and marketing each rebuild their own connections to the same systems.

That centralization is also a cost lever. Jain pointed to two mechanisms: model routing, which sends most tasks to smaller, cheaper models instead of defaulting to frontier models, and precomputed context, which avoids an agent burning tokens assembling its own context from scratch.

"Claude is also very slow because the first part of assembling that context actually takes forever," Jain said. Routing that request through Glean instead, he said, can cut token consumption by as much as half.

What Zillow and Glean's approach means for enterprises

Across data, cost and permissions, the session offered a few practical takeaways for enterprises building agentic AI on their own systems.

Build the measurement baseline before the AI push, not after. Roberts said Zillow's ability to credibly attribute a 40% increase in shipped code to AI adoption rests on a DORA metrics baseline the team put in place years earlier, not on the AI rollout itself.

Centralize context once instead of letting every team rebuild it. Jain's core argument for Glean's platform is that duplicated integration work across finance, legal and marketing teams is a hidden cost most enterprises haven't accounted for.

Don't assume permission inheritance is enough for regulated data. Even with a permissions-aware context platform in place, Zillow layered hard rules and a standing compliance check on top for its most sensitive categories, rather than trusting the architecture to handle it automatically.

Treat context as a cost lever, not just a capability. Model routing and precomputed context were the two mechanisms Jain pointed to for cutting AI spend, both aimed at reducing wasted token consumption rather than adding new capability.

"Models by themselves are not enough to bring automation with AI inside your enterprise," Jain said. "You do have to connect it with your enterprise context."