Why Anthropic Acquired Vercept for Computer Use

This analysis explains how vercept Changed Claude's Computer-Use Performance through the source's decisions, evidence and operating constraints.

Read the evidence below as a decision trail: what changed, why it matters, which trade-offs shaped the result, and where the conclusion still depends on context.

Anthropic acquired Vercept, an AI perception and interaction technology company, and three co-founders joined the team. Claude's computer operation ability jumped from less than 15% at the end of 2024 to 72.5% on the OSWorld benchmark test, approaching human levels. This is Anthropic's second major team acquisition after acquiring Bun.

[Core Insight] Anthropic uses mergers and acquisitions to complement the “eyes and hands” of AI Agents

AI can write codes and answer questions, but in order for it to operate a computer like a human—understanding the buttons on the screen, switching fields in a spreadsheet, and filling out forms across tabs—what needs to be solved is the problem of "perception and interaction." Anthropic announced the acquisition of Vercept on February 25, 2026, precisely to complement this capability.

Vercept’s team has spent years focusing on how AI systems “see and act” in software that humans use every day. Three co-founders Kiana Ehsani, Luca Weihs and Ross Girshick will join Anthropic to directly promote Claude's computer use capabilities. Vercept will end its external product operations in the coming weeks.

【In-depth dismantling】

What does the OSWorld benchmark jump from less than 15% to 72.5% mean?

OSWorld is a widely used benchmark for evaluating AI computer performance. At the end of 2024, the Sonnet model scored less than 15% on this benchmark — meaning the AI ​​would fail on most real computer operations tasks.

So far, Sonnet 4.6 has reached72.5%. Anthropic said Claude is now approaching human-level performance on tasks such as "navigating complex spreadsheets" and "complete web forms across browser paginations." From 15% to 72.5%, an increase of nearly 5 times in about 14 months. This rate of improvement shows that computer operations are no longer just experimental functions, but are becoming usable productivity tools.

Why did Anthropic choose to acquire rather than build in-house?

The core problem of letting AI operate computers is not just language understanding;Visual perception and interaction— The AI must understand the meaning, location, and operation of every element on the screen. This is a specialized field that requires many years of in-depth study. The founding premise of Vercept is that "to make AI truly useful in completing complex tasks, we need to solve difficult perception and interaction problems."

This is Anthropic’s second major team acquisition, following the acquisition of Bun. Anthropic said in the announcement that they give priority to teams whose technical ambitions are consistent with their own and who value security and rigorous principles. This strategy of "acquiring teams rather than products" shows that Anthropic believes that Computer Use is a core competency that requires long-term investment, rather than a function that can be quickly added.

What problems can Computer Use solve that "pure programming cannot solve"?

Anthropic notes that Claude's computer skills allow it to "handle multi-step tasks in real applications, just like a human sitting at a keyboard" and "solve problems that cannot be solved with code alone." The key point of this sentence is: many enterprise processes do not have APIs, and employees complete their work through GUI interfaces (click, enter, switch pagination). Computer Use allows AI Agents to directly operate these interfaces without waiting for each software to develop API connections.

【Takeaways: The value you can take away】

1. The “perception layer” of AI Agent is the next focus of competition: Language skills have matured, and whether the AI Agent can "understand and operate" real software interfaces will determine whether the AI Agent can truly replace humans' repetitive computer work.
2. OSWorld 72.5% is a major milestone: From less than 15% to 72.5% in only about 14 months. This speed suggests that Computer Use may go from "usable" to "easy to use" within a year.
3. Software without API can also be operated by AI: The value of Computer Use lies in skipping the API connection and directly operating the GUI interface. For businesses using a lot of legacy software, this is the lowest-friction path to AI adoption.
4. M&A strategy reveals priorities: Anthropic has acquired teams (Bun, Vercept) twice in a row instead of acquiring products, which shows that they are making a long-term layout on Agent infrastructure rather than short-term functional stacking.

---

Original source:Anthropic acquires Vercept to advance Claude’s computer use capabilities

What to take away

The practical value lies in the decisions and constraints behind vercept Changed Claude's Computer-Use Performance, not in copying one implementation without its context.