Anthropic just paid $50 million for a Seattle startup that Meta had already picked apart. The acquisition of Vercept, announced this week, brings three AI2 researchers to Claude’s computer use team - after Meta offered a fourth co-founder a reported $250 million to join its Superintelligence Lab instead.
The deal tells two stories at once: the escalating war for AI talent and the race to build AI systems that can operate computers like humans do.
What Anthropic Bought
Vercept was founded in late 2024 by five researchers from the Allen Institute for AI (AI2): Matt Deitke, Kiana Ehsani, Ross Girshick, Luca Weihs, and Oren Etzioni. Etzioni founded AI2 itself. Girshick previously led computer vision research at Facebook AI Research.
The team built Vy, a cloud-based agent that could remotely operate macOS environments. Users gave natural language commands; Vy navigated apps, clicked buttons, filled forms, and completed multi-step workflows. It was, in effect, a remote-controlled Mac operated by AI.
The startup raised $50 million total, including a $16 million seed round in January 2025 from Seth Bannon’s Fifty Years VC, with participation from Eric Schmidt, Jeff Dean, Kyle Vogt, and Arash Ferdowsi.
Then things got complicated.
The Meta Poaching
Last summer, Meta CEO Mark Zuckerberg personally recruited co-founder Matt Deitke. According to reports, Meta offered $250 million over four years to bring the 24-year-old researcher to its Superintelligence Lab. Deitke accepted.
Losing a co-founder to a $250 million offer would strain any startup. For Vercept, it accelerated the decision to find a buyer.
Three co-founders - Ehsani, Weihs, and Girshick - are joining Anthropic to work on Claude’s computer use capabilities. Etzioni, who remains involved with AI2 and its incubator, is not joining. Deitke congratulated his former colleagues on X but stayed at Meta.
The Product Dies
Vy, the product Vercept built, shuts down March 25. Customers have 30 days to migrate.
This is standard for talent acquisitions: the company buys the team, not the product. Anthropic needs researchers who understand “hard perception and interaction problems” for computer use, not a competing agent platform.
Why Computer Use Matters
Claude’s computer use capabilities have improved dramatically. On OSWorld - a benchmark that measures AI’s ability to navigate GUIs, complete forms, and execute multi-step workflows - Claude Sonnet 4.6 scores 72.5%.
That’s near human performance. The benchmark’s human baseline is approximately 72%.
The trajectory shows how fast this is moving: Sonnet 3.5 scored 14.9% in late 2024. Sonnet 3.5 v2 hit 28.0%. Sonnet 3.6 reached 42.2%. Sonnet 4.5 hit 61.4%. Now Sonnet 4.6 is at 72.5%.
Sixteen months from 15% to human parity.
GPT-5.2, for comparison, scores 38.2% on the same benchmark. Anthropic has a substantial lead in this specific capability.
The Competition Responds
The same week Anthropic announced the Vercept acquisition:
Microsoft opened a waitlist for Copilot Tasks, a feature that accepts natural-language goals, creates a plan, and executes it using browser and app orchestration. It asks permission before spending money or sending messages.
Google rolled out Gemini task automation on Android. Users on Pixel 10 and Samsung Galaxy S26 can tell Gemini to book Uber rides or order food from DoorDash. The AI operates apps in a secure virtual window while you continue using your phone.
Meta continues building its Superintelligence Lab, where Deitke now works on “multimodal superintelligence” - which includes computer use as a core capability.
Everyone is racing toward the same goal: AI that can operate software the way humans do.
UiPath Feels the Heat
When Anthropic announced the acquisition, UiPath’s stock dropped 3.6%.
UiPath built its business on robotic process automation (RPA) - software that automates repetitive computer tasks through scripted workflows. The company has spent the past year pivoting toward “agentic AI orchestration,” trying to stay relevant as AI models learn to do what RPA bots have always done.
The problem: AI computer use doesn’t need scripted workflows. It adapts. It figures out how to navigate an unfamiliar interface. It handles the edge cases that break traditional automation.
UiPath joined the Agentic AI Foundation as a Gold Member on February 23. Two days later, Anthropic acquired the team that might make UiPath’s core product obsolete.
Anthropic’s Acquisition Pattern
Vercept is Anthropic’s second major acquisition in three months. In December, the company acquired Bun, a JavaScript runtime and coding engine.
The pattern is clear: acquire teams that accelerate specific capabilities. Bun improved Claude’s coding infrastructure. Vercept advances computer use. Both acquisitions bring researchers who’ve spent years on hard problems Anthropic needs solved.
This is expensive. The Vercept deal reportedly valued the startup at $50 million despite having no significant revenue. Anthropic is paying for talent and expertise, not business fundamentals.
What This Means
Computer use is becoming a core battleground in AI development. The ability to operate software - not just generate text about software - creates enormous value. AI that can book flights, manage emails, navigate enterprise applications, and complete complex workflows changes what’s possible.
The talent war is intensifying. When Meta offers $250 million to poach a single researcher, when Anthropic acquires an entire startup for its team, when Google and Microsoft race to ship computer use features - this is what it looks like when capability becomes existential.
For users running local AI or considering enterprise deployments, the competitive dynamics matter. Claude’s computer use lead is substantial today. Whether it holds depends on how quickly Microsoft, Google, and Meta close the gap - and whether Anthropic can integrate the Vercept team fast enough to stay ahead.
The Bottom Line
Anthropic acquired Vercept’s team after Meta poached a co-founder. The product dies; the researchers join Claude’s computer use effort. With AI computer use approaching human parity on standard benchmarks, every major AI lab is racing to build systems that can operate software autonomously. The talent acquisition game is getting expensive. The capabilities are getting real.