Anthropic acquired Seattle AI startup Vercept on February 25, just months after Meta poached one of Vercept’s four co-founders with a $250 million four-year contract. The deal brings three remaining founders—Kiana Ehsani, Ross Girshick, and Luca Weihs—to Anthropic to advance Claude’s “computer use” capabilities: AI that navigates desktop applications like a human with a laptop. Vercept’s desktop automation product Vy will shut down within 30 days.
The acquisition exposes the brutal economics of the AI talent war. Meta paid $250 million for one engineer. Anthropic likely acquired the entire remaining team for less than that single poaching package.
The $250M Poaching That Triggered an Acqui-Hire
Meta paid Matt Deitke, a 24-year-old AI researcher and Vercept co-founder, $250 million over four years to join Meta’s Superintelligence Lab in mid-2025. After a personal meeting with Mark Zuckerberg, the offer doubled from an initial $125 million. Six months later, Anthropic acquired the remaining Vercept team for an undisclosed sum almost certainly less than Deitke’s individual package.
This isn’t an isolated case. OpenAI averages $1.5 million in stock compensation per employee, with research scientists earning base salaries up to $685,000. Meta is offering $300 million packages to top engineers across the industry. Sam Altman confirmed Meta tried to lure OpenAI talent with $100 million signing bonuses.
Vercept’s trajectory reveals the new AI startup endgame. Founded in November 2024, the company raised $50 million by January 2025 (including a $16 million seed round valued at $67 million post-money). Investors included Eric Schmidt, Jeff Dean, Kyle Vogt, and Arash Ferdowsi. Eighteen months later, it’s an Anthropic acqui-hire. The traditional path was build → grow → IPO. The new path: build → lose founders to $250M offers → sell remaining team before the wheels fall off.
What “Computer Use” AI Actually Means
“Computer use” AI refers to systems that see screens, move cursors, click buttons, and type text—controlling desktop applications like a human with a laptop. Unlike chatbots that answer questions, computer use AI takes action: filling forms, navigating spreadsheets, automating workflows across multiple apps without APIs or scripting.
Vercept’s Vy product demonstrated the capability. Users instructed Vy via natural language (“export this Excel data to Slack”) or used “watch and repeat” mode where Vy recorded actions and learned workflows. The app handled cross-app automation, scheduled tasks, and data passing between applications. It will shut down in late March 2026, leaving users to migrate to Claude’s computer use feature or traditional RPA tools.
The technology is advancing fast. Anthropic’s Claude improved from under 15 percent on the OSWorld benchmark in late 2024 to 72.5 percent in early 2026—approaching human-level performance on complex spreadsheet navigation and multi-tab web forms. That’s a 4.8x improvement in 18 months. However, a 72.5 percent success rate still means the AI fails roughly one in four times on complex tasks. Good enough for augmentation with human oversight, not fully autonomous operation.
Related: Gemini Android Automation: AI Agents Order Food & Rides
The Public Investor Fight on LinkedIn
The acquisition sparked a rare public fight between Vercept investors. Co-founder Oren Etzioni, founder of the Allen Institute for AI, called the deal “throwing in the towel” on LinkedIn and blamed lead investor Seth Bannon for not hiring correct business leadership. Vercept had technical traction, Etzioni argued, but exited too early.
Bannon fired back: “You disparaged the heroic work of the founders for achieving an outcome most could only dream of.” The two traded accusations of lies and legal threats publicly—unusual behavior for sophisticated investors who typically keep post-acquisition disputes private. The fight revealed deep disagreement about strategy: was this a successful outcome (investors got returns) or a premature exit (“throwing in the towel” after just 18 months)?
The answer depends on perspective. Losing Matt Deitke to Meta’s $250 million offer likely destabilized Vercept. Building a company with one founder suddenly at a competing lab is difficult. An acqui-hire to Anthropic provided liquidity for investors and a path forward for the remaining team. But it’s hard to call it a success when the company raised $50 million to become a feature in someone else’s product.
Why Anthropic Wants This: The Computer Use Race
Anthropic is racing OpenAI, Google, and Meta to dominate computer use AI. Claude was the first frontier model to launch the capability in October 2024, but competitors are moving fast. Google announced Gemini Android automation on February 25—the same day as the Vercept news—enabling AI to order Uber rides and Doordash deliveries. OpenAI is building coding agents through its Codex app. Meta is stockpiling talent (like Matt Deitke) for its Superintelligence Lab.
The stakes are high. Gartner predicts 40 percent of enterprise applications will embed AI agents by the end of 2026, up from less than 5 percent in 2025. The agentic AI market is projected to grow from $7.8 billion today to $52 billion by 2030. Anthropic currently trails OpenAI in enterprise adoption—77 percent versus 40 percent in recent surveys. Acquiring Vercept’s expertise helps close that gap.
This is Anthropic’s second acquisition in three months. In December 2025, it acquired the Bun team (a JavaScript runtime and coding engine). The pattern is clear: acquire teams with “strong technical ambitions” that “advance core capabilities.” Computer use is a core capability. Vercept’s team brings deep expertise in visual perception and interaction within software applications—exactly what Anthropic needs to maintain its lead.
What Happens Next
Vercept’s Vy desktop app will shut down within 30 days. Users need to migrate to Claude’s computer use feature, traditional RPA tools like UiPath, or manual processes. Anthropic hasn’t disclosed integration timelines or specific features, only that Vercept’s team will “advance Claude’s computer use capabilities.” Matt Deitke remains at Meta and congratulated his former colleagues on X, confirming he’s not joining Anthropic.
The talent war shows no signs of slowing. Meta is paying $250 million for individual engineers. Anthropic is responding by acquiring entire teams. OpenAI is defending with $1.5 million average stock packages. The computer use AI race is accelerating as the industry pivots from chatbots to action-taking agents.
For Vercept, it’s an unusual end: raise $50 million, build promising technology, lose a founder to a quarter-billion-dollar offer, exit to a competitor in 18 months. Whether that’s success or failure depends on who you ask—and whether you’re Oren Etzioni or Seth Bannon.
