On February 25, 2026, Anthropic announced the acquisition of Vercept, a Seattle-based AI startup that had spent two years building tools for AI agents to see and operate software the way a person at a keyboard does. The deal is Anthropic's second-ever acquisition. Its stated purpose is straightforward: push Claude's computer-use capabilities further, at a moment when the benchmark scores are close enough to human performance that the remaining distance is worth closing aggressively.
Vercept was founded by Kiana Ehsani, Luca Weihs, and Ross Girshick, three researchers with backgrounds in computer vision and embodied AI. The company's flagship product was Vy, a cloud-based agent that could operate a remote Apple MacBook, navigating live applications, filling forms, and completing multi-step workflows without human supervision. Vy shut down on March 25, 2026, as the Vercept team transitioned fully inside Anthropic. The deal was reportedly valued at around $50 million, making it primarily a talent acquisition.
From a Standing Start to 72.5%
When Anthropic first introduced computer use in late 2024, Claude Sonnet models scored below 15% on OSWorld, a benchmark designed to measure how well AI agents complete realistic tasks inside real desktop applications. OSWorld is deliberately hard: it tests agents against ordinary software that millions of people use every day, across a range of operating systems and interface types. A 15% score meant the early versions of Claude computer use failed on roughly five out of six attempts.
By the time the Vercept deal was announced in February 2026, that score had reached 72.5%. Claude Sonnet 4.6 can now navigate spreadsheets, complete multi-step web forms across multiple browser tabs, and extract structured information from desktop applications that have no API. Human-level performance on OSWorld sits in the 75-80% range, depending on the task category, so the gap has narrowed from enormous to marginal.
That shift matters because the usefulness of a computer-use agent changes qualitatively around the 70% range, not just quantitatively. Below 30%, agents fail often enough that users need to monitor and correct continuously, which eliminates most of the efficiency gain. At 70% and above, agents can handle a broad enough range of tasks autonomously that they become genuinely useful for production workflows, not just demos.
Vercept Acquisition: Key Facts
- Acquisition announcedFebruary 25, 2026
- Vercept productVy, a cloud-based computer-use agent
- Vy shutdown dateMarch 25, 2026
- OSWorld score at launch (late 2024)Under 15%
- OSWorld score at acquisition72.5%
- Vercept co-foundersKiana Ehsani, Luca Weihs, Ross Girshick
Why Vercept, and Why Now
The Vercept founders built their company around a specific idea: making AI genuinely useful for complex tasks requires solving hard perception problems, not just language ones. Generating accurate text is one skill. Correctly identifying a dialog box in an unfamiliar application, clicking the right option, waiting for the result, and then deciding what to do next is a different skill entirely. The gap between those two capabilities has been one of the sharper dividing lines between what AI can currently do and what enterprise customers actually need.
Computer use fills that gap by giving Claude a path through any software interface, regardless of whether an API exists. That matters a great deal in enterprises where critical data lives in legacy systems that were built for human operators and never updated for programmatic access. It also matters for individual professionals who need help inside the specific applications they use every day, from older ERP systems to proprietary industry tools that will never offer a developer API.
Ross Girshick, one of Vercept's co-founders, built his career at the intersection of visual perception and action, work that now maps directly onto the problems Anthropic needs to solve. The acquisition is a talent bet as much as a technology one: the perception and interaction challenges involved in reliable computer use require researchers who have spent years on nothing else, and that concentration of expertise is not easily replicated by general-purpose AI research teams.
"The Vercept team has spent years thinking carefully about how AI systems can see and act within the same software humans use every day." Anthropic, acquisition announcement, February 2026
How This Fits Anthropic's Acquisition Strategy
The Vercept deal follows Anthropic's acquisition of Stainless in May 2026, a deal aimed at strengthening the SDK infrastructure developers use to build on top of Claude. Where Stainless addressed the integration layer for developers working with well-structured APIs, Vercept extends the stack in a different direction: the ability to operate inside applications that have no SDK, no API, and no programmatic interface at all. Together, the two acquisitions address two of the most common reasons that Claude agent deployments stall in real enterprise environments.
The timing also connects to the broader agentic infrastructure Anthropic has been building. Managed agents, proactive workflows, and multi-agent orchestration all depend on reliable environment interaction, not just reliable language generation. An agent that can navigate a live application adds a capability dimension that conversation-only agents cannot match, and Anthropic's product roadmap for 2026 has been built around that assumption.
Whether the Vercept team's work accelerates Claude's computer-use reliability on the timeline Anthropic's enterprise customers expect is a question that will be answered over the next two quarters. The benchmark scores suggest the foundation is solid. The harder problem is performance across the specific, messy, inconsistent software environments that real enterprise customers actually run, which are considerably more varied than any benchmark can capture. That is exactly the problem the Vercept team was hired to solve.
