Anthropic buys Vercept to master computer use

Giving AI the ability to use a computer just like a person at a keyboard is turning out to be one of tech's fiercest battlegrounds.

Anthropic announced on Wednesday that it has acquired Vercept, a Seattle-based startup focused on AI perception and interface interaction. The deal marks Anthropic's second acquisition in recent months, following its purchase of coding runtime engine Bun in December.

As part of the deal, Vercept will wind down its commercial product by March 25, giving users just 30 days to pack up. Key members of the team are heading straight into Anthropic to supercharge Claude's computer use features.

Teaching agents to run a MacBook

Vercept spent the past year building systems that help AI models see, understand, and interact with human software. Its flagship cloud agent, named Vy, was designed to remotely pilot an Apple MacBook to perform multi-step administrative workflows.

The startup had impressive pedigree, spinning out of the Seattle-based Allen Institute for AI (AI2). It had raised roughly $50 million in total funding, including a $16 million seed round early last year, backed by angel heavyweights like former Google CEO Eric Schmidt, Google DeepMind chief scientist Jeff Dean, Cruise founder Kyle Vogt, and Dropbox co-founder Arash Ferdowsi.

Anthropic says Vercept's research maps directly onto Claude's roadmap. Just recently, Anthropic rolled out Claude Sonnet 4.6, showcasing dramatic leaps on the OSWorld benchmark. The model hit 72.5%, compared to under 15% when computer use first debuted in late 2024, approaching human proficiency on tasks like cross-tab web forms and intricate spreadsheets.

Star talent and a messy investor feud

While CEO Kiana Ehsani and co-founders Luca Weihs and Ross Girshick are making the jump to Anthropic, not everyone from the founding circle is coming along.

One co-founder, Matt Deitke, made headlines last year after reportedly landing an eye-popping $250 million compensation package to join Meta's Superintelligence Lab. He took to X on Wednesday to congratulate his former teammates.

Meanwhile, prominent Seattle technologist and founding AI2 director Oren Etzioni, who was an early co-founder and investor in Vercept, stayed behind and made his frustration public. On LinkedIn, Etzioni lamented that after barely a year and strong momentum, the company was essentially throwing in the towel and cutting off customers. While he noted he made a return on his investment, he called the quick exit sad.

That sparked an unusually heated public spat. Seth Bannon, a partner at lead investor Fifty Years, pushed back hard in the comments, accusing Etzioni of disparaging the founders' achievements. The exchange quickly escalated into accusations of misrepresentation and legal threats before the thread quieted down.

Ehsani framed the move as the most efficient path forward, posting that the team could either keep building alone toward the same vision or join Anthropic to accelerate it into reality.

Why it matters

We think this deal highlights where the frontier AI race is heading next: desktop autonomy.

For builders and end users, Anthropic's playbook is becoming remarkably clear. Buying Bun gave Claude Code a turbocharged foundation for running local code; acquiring Vercept brings the vision and interaction smarts required to operate graphical user interfaces. If Anthropic glues these pieces together seamlessly, Claude moves from an assistant that merely suggests code to an operator that can independently audit spreadsheets, enter enterprise data, and bridge legacy software. The harsh trade-off, however, is reliability for early adopters. Vercept's 30-day shutdown is a stark reminder that relying on point-solution agent startups carries serious rug-pull risk.

For the broader industry, the acquisition shows how narrow the window has become for independent agent startups. Even with $50 million in funding and an enviable bench of AI royalty on the cap table, sustaining a standalone wrapper around computer interaction proved difficult to justify against foundation model labs. Meta snatched one founder with historic compensation, and Anthropic absorbed the rest.

Anthropic played this smartly by aggressively securing top-tier computer vision talent before competitors could. But the public discord among investors underscores real friction in how fast early-stage AI ventures are being folded into hyper-scalers. Claude's computer use is advancing fast on paper, but the true test will be whether this newly consolidated team can turn synthetic OSWorld benchmarks into dependable daily automation.