Quick Brief
- Anthropic acquired Seattle AI startup Vercept on February 25, 2026, bringing its nine-person team into Claude’s development
- Claude Sonnet’s OSWorld score rose from under 15% in late 2024 to 72.5% today, approaching human-level task performance
- Vercept co-founders Kiana Ehsani, Luca Weihs, and Ross Girshick bring deep AI perception and vision expertise to Anthropic
- Vercept’s external product shuts down within weeks as the full team integrates into Anthropic’s computer use roadmap
AI just crossed a threshold that quietly changes how digital work gets done. Anthropic’s acquisition of Vercept is not a talent grab or a defensive move. It is a direct investment in making Claude the most capable computer-using AI agent available. The bottleneck has always been perception, and this team was built specifically to solve it.
Why Vercept Was the Right Acquisition for Anthropic
Vercept was built around one clear thesis: making AI genuinely useful for completing complex tasks requires solving hard perception and interaction problems. Founded by alumni of the Allen Institute for AI (Ai2), the Seattle startup was not a generalist lab. It was a focused research team attacking one of the hardest unsolved problems in the field: how to make AI perceive and act inside graphical software the same way a person at a keyboard does.
Co-founder Kiana Ehsani brings years of research into AI perception and embodied systems from her work at Ai2. Co-founder Ross Girshick is a recognized expert in vision AI, with deep roots in computer vision research from his time at Facebook AI Research (FAIR) and Microsoft Research. Co-founder Luca Weihs specializes in AI agents operating in complex, real-world-like software environments. That combination maps directly onto the hardest problems Anthropic is working on with Claude’s computer use capability.
The nine-person team joins Anthropic in full, and Vercept’s external product winds down in the coming weeks.
Perplexity Computer Is the General-Purpose AI Worker That Handles Entire Projects, Not Just Prompts
Claude’s OSWorld Score Jumped from Under 15% to 72.5% in One Year
The numbers define the progress precisely. When Anthropic first released computer use in late 2024, Claude’s score on OSWorld, the widely-used benchmark for AI computer interaction, sat below 15%. With the recent launch of Claude Sonnet 4.6, that figure reached 72.5%.
OSWorld tests real-world computer tasks: navigating complex spreadsheets, completing multi-step web forms across browser tabs, and coordinating actions inside live applications. A 72.5% score means Claude handles the vast majority of these tasks at a level approaching human performance. The Vercept acquisition is designed to push those numbers further and close the remaining gap.
What Computer Use Actually Means in Practice
Computer use is not screen-sharing or remote desktop access. It means Claude can open applications, click buttons, fill forms, read on-screen content, and complete multi-step workflows inside live software, the same way a person at a keyboard would, with no API integration required from the target application.
That distinction matters for real work. Traditional automation approaches depend on rigid scripts tied to specific interface structures, which break when software updates. Claude’s vision-based approach reads the screen contextually and adapts, making it applicable to legacy systems, web portals, and standard desktop environments that have no special AI integration. Enterprises dealing with data-heavy, multi-step processes, such as invoice handling, form-based compliance workflows, and cross-tool research tasks, are the most immediate beneficiaries.
We tested Claude Sonnet 4.6’s computer use over 10 days across 6 workflow categories including spreadsheet navigation, browser form completion, and cross-application data transfer. Performance aligned closely with Anthropic’s reported benchmark improvements, with the strongest results appearing in structured, predictable UI environments.
The Perception Problem Vercept Was Built to Solve
Most assumptions about AI difficulty center on language. For computer use, the real challenge is vision. An AI agent must identify interactive elements on a screen, understand their context, predict the result of engaging with them, and recover from unexpected states without human intervention. That requires a fundamentally different research focus than text generation.
Vercept’s team spent years at Ai2 developing tools for embodied AI and vision-based interaction in complex environments. That prior research is a direct technical predecessor to the perception infrastructure Claude needs to operate reliably inside dynamic, real-world software. Bringing that expertise in-house, rather than trying to replicate it, is why this acquisition is strategically significant.
Anthropic’s Acquisition Strategy: Targeted Technical Expertise
Vercept is the second team Anthropic has brought in, following the acquisition of Bun. The pattern is consistent: Anthropic acquires small, technically focused teams working on specific capability challenges, not large companies with established revenue or market share. Vercept had nine people, raised a $16 million seed round led by Fifty Years with participation from Eric Schmidt and Jeff Dean, and built a focused product before joining Anthropic.
This approach reflects a deliberate strategy to accelerate capability development in areas where internal progress alone would be slower. With competitors including OpenAI and Google DeepMind also investing heavily in agentic AI systems in 2026, targeted acquisitions of expert teams represent an efficient path to closing technical gaps.
What This Means for Developers and Enterprise Users
If you build on Claude’s API, the Vercept integration signals that computer use reliability, precision, and task coverage will continue improving materially over coming months. Current deployments using Claude for computer use should anticipate expanded capability across more complex and dynamic software environments.
For non-technical users, the practical impact is direct: Claude will handle more of the repetitive digital work that still consumes significant time, from moving data between systems to navigating multi-step web portals, with less manual oversight required as accuracy continues to rise toward and past the current 72.5% benchmark.
Limitations and Considerations
At 72.5% OSWorld accuracy, Claude Sonnet 4.6 approaches human-level performance on benchmark tasks, but the remaining gap is real. Dynamic interfaces, error-state recovery, and highly variable UI environments represent scenarios where human oversight remains necessary. Anthropic’s own guidance recommends careful, supervised deployment for sensitive or irreversible actions until accuracy improves further.
Claude Under Attack: How Three Chinese AI Labs Extracted 16 Million Exchanges from Anthropic
Frequently Asked Questions (FAQs)
What is the Anthropic Vercept acquisition?
Anthropic acquired Vercept, a Seattle AI startup specializing in vision-based computer perception and automation, on February 25, 2026. The nine-person team, including co-founders Kiana Ehsani, Luca Weihs, and Ross Girshick, joins Anthropic to advance Claude’s ability to operate live software like a human user.
Who founded Vercept and what is their background?
Vercept was co-founded by Kiana Ehsani, Luca Weihs, and Ross Girshick, all with deep roots in AI research at the Allen Institute for AI (Ai2). Ross Girshick is a recognized computer vision expert with prior work at Facebook AI Research (FAIR) and Microsoft Research. The team focused specifically on AI perception and interaction with real-world software environments.
What is OSWorld and why does Claude’s 72.5% score matter?
OSWorld is a widely-used benchmark that tests AI models on real computer tasks including navigating spreadsheets and completing multi-step web forms across browser tabs. Claude Sonnet 4.6’s score of 72.5%, up from under 15% in late 2024, represents near-human-level performance on the benchmark and the strongest result Anthropic has publicly reported for computer use.
How does Claude’s computer use work without an API integration?
Claude uses vision-based perception to read live screen content, identify interactive elements, and execute actions including clicks, form inputs, and multi-step navigation. It does not require the target application to provide a special API or integration, making it applicable to legacy systems, standard web applications, and any desktop environment a human can operate.
Will Vercept’s product continue after the acquisition?
No. Vercept is winding down its external product, a Mac application called Vy that used vision to control computers, in the weeks following the acquisition. The entire team is integrating into Anthropic exclusively to advance Claude’s computer use capabilities.
What did Anthropic acquire before Vercept?
Anthropic’s previous acquisition was Bun. Both acquisitions reflect Anthropic’s stated strategy of bringing in small, technically elite teams whose work directly advances specific capability or infrastructure goals.
How much did Vercept raise before being acquired?
Vercept raised $16 million in a seed funding round in June 2025, led by Fifty Years. Notable investors included Eric Schmidt and Jeff Dean. The financial terms of the Anthropic acquisition have not been disclosed.
What types of tasks can Claude now handle with computer use?
Claude Sonnet 4.6 can navigate complex spreadsheets, complete web forms across multiple browser tabs, and manage multi-step workflows inside live applications without any special integration from those applications. Anthropic also notes broader use cases including writing and running code across repositories and synthesizing research from multiple tools and sources.
