Anthropic acquires Vercept: what it means for Claude's computer-use future
Anthropic bought Vercept, an AI startup that built cloud agents operating full desktops. Here is why this acquisition matters.
Whether you're looking for an angel investor, a growth advisor, or just want to connect — I'm always open to great ideas.
Get in TouchAI, startups & growth insights. No spam.
Anthropic has acquired Vercept, a Seattle-based AI startup that built Vy, a cloud agent capable of operating a full remote MacBook on your behalf. The deal, announced February 25, 2026, is Anthropic's second-ever acquisition and signals a serious push to make Claude the go-to AI for computer automation. Vercept's product shuts down March 25, and the entire founding team, except one, is joining Anthropic.
Vercept's flagship product was Vy, a computer-use agent that lived in the cloud and operated a remote Apple MacBook. You gave it a task, and it navigated your screen just like a person would: clicking, typing, scrolling, filling forms, switching between apps.
The key distinction from older robotic process automation (RPA) tools is how it "saw" the screen. Traditional automation tools like UiPath connect to application APIs or rely on pre-written scripts that break the moment a UI changes. Per Anthropic's announcement, Vy used computer vision to interpret whatever was on the screen at runtime, no API needed, no brittle selectors. That approach made it flexible enough to work across basically any desktop application.
The startup was founded in November 2024 by Kiana Ehsani (CEO), Luca Weihs, Ross Girshick, and Matt Deitke. The team came with serious research credentials. Girshick is one of the original authors of the R-CNN family of object detection models, foundational work in modern computer vision. Deitke had received an Outstanding Paper Award at NeurIPS 2022 for work on 3D datasets and embodied AI. Weihs had deep experience building AI research environments.
Vercept wasn't building another chatbot wrapper. It was going after the hard problem of making an AI agent reliable enough to operate real software in real conditions, with all the edge cases, pop-ups, slow load times, and broken UIs that entails.
Vercept raised $50 million in total funding, including a $16.5 million seed round in January 2025 that valued the company at $67 million post-money. Financial terms of Anthropic's acquisition were not disclosed.
As GeekWire reported, the investor list reads like a Silicon Valley all-star roster: former Google CEO Eric Schmidt, Google DeepMind chief scientist Jeff Dean, Cruise founder Kyle Vogt, and Dropbox co-founder Arash Ferdowsi all backed the startup. That caliber of angel investor, combined with the speed of the exit (roughly 15 months from founding to acquisition), says a lot about both the promise of the team and the compressed timelines AI startups now operate under.
Vercept's product, Vy, will officially shut down on March 25, 2026. Existing customers are being wound down with 30 days notice. The founding team, excluding Matt Deitke, is joining Anthropic.
Not everyone is thrilled. As TechCrunch noted, Oren Etzioni, an AI researcher and Vercept investor, described the outcome as "sad" on LinkedIn, writing he was "pleased to have gotten a positive return but obviously disappointed that after just a little over a year with so much traction, and such a fantastic team, we're basically throwing in the towel." Investor Seth Bannon responded with a quote from Theodore Roosevelt's "Man in the Arena" speech, a somewhat pointed response to critics second-guessing the decision.
Per SiliconAngle, CEO Kiana Ehsani framed it as a mission-driven choice: "The choices were clear: we could build independently and work toward the same vision as two separate versions of it, or join forces with an incredible team and accelerate that vision into reality. The decision became an easy choice."
Computer use is Anthropic's most ambitious product bet. Claude gained computer-use capabilities in October 2024 as a beta feature with Claude 3.5 Sonnet. At launch, it was impressive in demos and unreliable in practice. The OSWorld benchmark, which scores AI agents on their ability to complete real tasks inside real desktop applications, gave early Claude computer-use models a score below 15%.
That has changed dramatically. By early 2026, Claude Sonnet 4.6 scores 72.5% on OSWorld, up from sub-15% in under two years. Claude Opus 4.6 scores 72.7%. For context, OpenAI's GPT-5.2 trails at 38.2% on the same benchmark.
Anthropic is winning the benchmark war on computer use right now. But benchmarks and real-world deployment are different things. Vercept's team spent 15 months building, testing, and refining an agent that customers actually used on live machines. That production experience, knowing where agents fail in practice, how to handle edge cases, how to make them reliable enough that a user will trust them with their actual work, is exactly what Anthropic needs to move from "impressive demo" to "tool you depend on."
This is also Anthropic's second acquisition in three months. In December 2025, it acquired Bun, the high-performance JavaScript runtime used heavily in Claude Code infrastructure. That deal was about scaling developer tooling. The Vercept deal is about scaling AI agency itself.
The pattern is clear: Anthropic is moving from a lab that builds models to a company that ships products for real work.
Computer-use AI agents are now a serious competitive front. Here is where the major players stand as of February 2026.
| Agent / product | Provider | OSWorld score | Approach |
|---|---|---|---|
| Claude Opus 4.6 | Anthropic | 72.7% | Computer vision + model reasoning |
| Claude Sonnet 4.6 | Anthropic | 72.5% | Computer vision + model reasoning |
| GPT-5.3-Codex | OpenAI | 64.7% | Computer vision + model reasoning |
| GPT-5.2 | OpenAI | 38.2% | Computer vision + model reasoning |
| Gemini 3 Pro | Below GPT-5.2 | Computer vision + model reasoning | |
| UiPath | UiPath | N/A | Script-based RPA, no LLM core |
| Traditional RPA | Various | N/A | API connectors, brittle selectors |
The Vercept acquisition also rattled UiPath's stock. According to Investing.com, shares of UiPath (NYSE: PATH) fell 3.6% on the day of the announcement. The market read this correctly: an Anthropic-backed computer-use agent that can navigate any application without scripts or connectors is a direct threat to the RPA category that UiPath built its business on.
The fundamental difference between Vercept's approach (and where Claude computer-use is heading) versus traditional RPA is the absence of integration work. Setting up a UiPath workflow for a new application requires defining selectors, handling exceptions, maintaining scripts as UIs change. An AI agent that just looks at the screen and decides what to do next requires none of that upfront work. It is slower and more expensive per task today, but the setup cost is near zero.
One Vercept co-founder is not joining Anthropic. Matt Deitke, age 24, left Vercept earlier to join Meta's Superintelligence Lab on a reported $250 million package across four years, with up to $100 million payable in year one. Meta CEO Mark Zuckerberg personally met with Deitke before the offer was finalized.
This is the backdrop against which the Vercept acquisition happened. A major AI lab picked off one of the co-founders of a well-funded startup with a $250 million compensation package. That is not a normal recruiting outcome. It is a signal of how acutely the large labs feel the shortage of people who can actually build reliable agentic AI systems.
When Anthropic announced the acquisition, Deitke tweeted:
"Congrats to @ehsanik, @LucaWeihs, and @inkynumbers! Happy for you all" -- @mattdeitke
Publicly warm. But the subtext is that Meta got one of the four co-founders, and Deitke was, by multiple reports, not enthusiastic about the acqui-hire path.
This acquisition also shows Anthropic's counter-move in the talent war. Rather than matching $250 million compensation packages for individual researchers, Anthropic acqui-hired the rest of the team, getting three experienced founders with production agentic AI experience and absorbing their institutional knowledge. The cost of the acquisition was not disclosed, but it was almost certainly less than the $250 million Meta paid for one person.
The talent dynamic in agentic AI right now: there are a few dozen people in the world who have actually shipped agents that work reliably on real computers. Meta, Anthropic, Google, and OpenAI all want them. The competition for that talent is happening through compensation packages that would have seemed absurd two years ago.
If you use Claude's computer-use feature or are evaluating it for your team, here is the practical read.
According to SiliconAngle's analysis, Vercept's technology and team being inside Anthropic should accelerate Claude's ability to handle multi-step tasks inside real applications. The specific areas where Claude computer-use has struggled historically include: recovering gracefully from unexpected popups or UI changes, knowing when to stop and ask for confirmation versus proceeding autonomously, handling slow-loading applications without timing out, and maintaining context across long task sequences.
These are exactly the categories where a team that built and operated a production computer-use agent for 15 months would have hard-won experience.
The shutdown of Vy on March 25 is a loss for the small number of businesses that built workflows on it. But the 30-day notice gives them time to prepare, and the most likely outcome is that Claude's computer-use capabilities, now enriched by Vercept's engineering team, become a better long-term home for those use cases anyway.
For the broader ecosystem: this acquisition, combined with the UiPath stock reaction, is an early signal that the AI agent wave is not just theoretical. Companies that built businesses on scripted automation should be watching this space closely.
Vercept is a Seattle-based AI startup founded in November 2024 that built Vy, a cloud-hosted agent that could operate a full remote MacBook on a user's behalf. Rather than using scripts or API connectors like traditional automation tools, Vy used computer vision to navigate applications the way a human would.
Financial terms of the acquisition were not disclosed. Vercept had raised a total of $50 million before the deal, including a $16.5 million seed round in January 2025 that valued the company at approximately $67 million post-money.
Vercept was co-founded by Kiana Ehsani (CEO), Luca Weihs, Ross Girshick, and Matt Deitke. Ehsani, Weihs, and Girshick are joining Anthropic. Deitke departed earlier to join Meta's Superintelligence Lab.
Vercept's angel investors included former Google CEO Eric Schmidt, Google DeepMind chief scientist Jeff Dean, Cruise founder Kyle Vogt, and Dropbox co-founder Arash Ferdowsi.
Yes. Vercept's product Vy will shut down on March 25, 2026. Existing customers received 30 days notice following the acquisition announcement on February 25, 2026.
Matt Deitke joined Meta's Superintelligence Lab on a reported $250 million compensation package across four years. Meta CEO Mark Zuckerberg personally met with Deitke, age 24, before the deal was finalized. Deitke is one of the most cited young researchers in embodied AI and 3D datasets, with an Outstanding Paper Award from NeurIPS 2022.
As of February 2026, Claude leads on the OSWorld benchmark, the standard measure for AI agent computer-use performance. Claude Sonnet 4.6 scores 72.5% and Claude Opus 4.6 scores 72.7%. OpenAI's GPT-5.2 scores 38.2%, and GPT-5.3-Codex scores 64.7%. Early Claude computer-use models scored below 15% on this same benchmark in late 2024.
UiPath (NYSE: PATH) shares fell 3.6% on the day of the announcement. Investors interpreted Anthropic's acquisition as a sign that AI-native computer-use agents, which require no scripting or API connectors, could displace traditional robotic process automation (RPA) tools. Vercept's computer vision approach directly competes with the workflow automation market where UiPath operates.
No. Anthropic's first acquisition was Bun, the high-performance JavaScript runtime, announced in December 2025. That deal focused on scaling the infrastructure behind Claude Code, Anthropic's code generation tool. Vercept is the second.
Anthropic launched computer-use as a beta feature in October 2024 alongside Claude 3.5 Sonnet. At launch, it scored below 15% on the OSWorld benchmark. By early 2026, following continuous improvements and now the addition of Vercept's team, Claude scores above 72% on the same benchmark.
Legendary computer scientist Donald Knuth publicly confirmed that Claude Opus 4.6 solved an open mathematical conjecture he'd pursued for weeks, calling it 'a dramatic advance in automatic deduction.'
Anthropic releases Claude Opus 4 and Sonnet 4 with hybrid instant-and-extended thinking, setting new SWE-bench records at 72.5% and 72.7% respectively.
Check Point Research discloses two Claude Code vulnerabilities — CVE-2025-59536 (CVSS 8.7) for remote code execution and CVE-2026-21852 (CVSS 5.3) for API key theft — triggered by opening malicious git repos.