Anthropic acquires OpenAI to steal talent: What is the "Silicon Valley double threat" planning?

Beijing Time, February 26 — American AI leader Anthropic announced the acquisition of visual-driven computer automation AI startup Vercept, filling the visual gap in the Computer Use feature. This is Anthropic’s second acquisition after acquiring full-stack toolchain startup Bun in December 2025.

Computer Use is a core capability developed by Anthropic for its large model Claude, enabling direct AI control of computers. It allows Claude to “see screens, move the mouse, type on the keyboard, and operate software” like a human, completing multi-step, cross-application complex tasks. It was released alongside Claude 3.5 Sonnet in October 2024.

However, early in its development, the model faced bottlenecks in visual understanding accuracy, complex interface recognition, and dynamic scene processing, resulting in low operation success rates and task failures. Vercept’s expertise in high-precision UI recognition, spatial reasoning, dynamic interface tracking, and low-latency visual processing can help address these shortcomings.

According to Anthropic, developers are increasingly using Claude to handle complex tasks, including writing and running entire code repositories, synthesizing research from dozens of sources, and managing workflows across multiple tools and teams. Computer Use enables Claude to operate in real-time within various software, completing multi-step tasks that are difficult to accomplish with code alone. The acquisition of Vercept will further enhance these capabilities.

Vercept is headquartered in Seattle, USA, founded by alumni of the Allen Institute for AI (AI2). It positions itself as a vision-first AI agent, focusing on “screen reading and computer control” automation without APIs. The core team specializes in robotics, embodied intelligence, AI agents, and reinforcement learning.

Based on its technical background and founding team, Vercept completed a $16 million seed round in January 2025, with a valuation of approximately $67 million. Investors include former Google CEO Eric Schmidt, Google DeepMind Chief Scientist Jeff Dean, Cruise founder Kyle Vogt, among others. Vercept has raised a total of about $50 million subsequently.

Anthropic believes Vercept’s expertise in perception and interaction can be directly applied to some of the challenging problems it is working on. Vercept’s desktop application Vy will be shut down within 30 days. As a transition, Vercept encourages users to try Anthropic’s Claude tools as alternatives during the service discontinuation.

Regarding this acquisition, Vercept co-founder Luca Weihs stated that in the future, Claude’s ability to perform knowledge-based tasks will be comparable to its current coding capabilities. This could fundamentally change how humans interact with computers, making this interaction as important as the underlying model itself.

This acquisition reflects the increasingly fierce competition in developing AI agents capable of controlling computers and other devices to complete user tasks. Besides Anthropic, OpenAI is also actively involved. In mid-February, OpenClaw founder Peter Steinberger announced he would join OpenAI full-time to develop the next-generation personal AI agent. OpenClaw was not acquired or integrated into OpenAI but transitioned to an independent open-source foundation, maintaining MIT licensing and community governance. OpenAI provides funding, technology, and model resources but does not own or control the project. This talent acquisition strategy aims to strengthen OpenAI’s capabilities in personal AI agents and counterbalance Anthropic’s lead in enterprise-level agents.

Although their approaches differ, both companies are betting on the AI agent track. Recently, Dario Amodei, founder of Anthropic, stated in an interview that the ultimate form of AI evolution is not text generation but a closed-loop execution system with computer control (Computer Use). It can observe screens, review history, understand context, and collaborate across platforms to perform complex tasks like video editing and software engineering, similar to human capabilities.

He believes the main obstacle in the industry is not intelligence but interaction reliability—meaning the model’s ability is sufficient, but reliably and accurately completing tasks without errors still needs improvement. In the next one to three years, AI will combine long-context management with computer vision control to address most jobs requiring “on-the-job learning.”

(Source: Yicai)