Google DeepMind today released the Gemini 2.5 Computer Use model in public preview via the Gemini API on Google AI Studio and Vertex AI. The specialized model, built on Gemini 2.5 Pro, powers AI agents that can interact with graphical user interfaces by clicking, typing, scrolling, and manipulating elements like dropdowns and filters. It operates in a loop: developers provide a screenshot, task description, and action history; the model returns function calls that client-side code executes, then a new screenshot is fed back until the task is done.

The model is optimized for web browsers and shows promise for mobile UI control, but is not yet optimized for desktop OS-level control. Google claims it outperforms leading alternatives on multiple web and mobile benchmarks, including Online-Mind2Web, WebVoyager, and AndroidWorld, with lower latency. Evaluations from Browserbase and Google’s own tests show around 70% accuracy on Online-Mind2Web. Safety features include a per-step safety service and developer controls to require user confirmation for high-stakes actions like purchases.

Early testers have used the model for personal assistants, workflow automation, and UI testing. Google teams have deployed it in production for UI testing and to power Project Mariner, the Firebase Testing Agent, and agentic capabilities in AI Mode in Search. The launch follows OpenAI’s Dev Day announcements and continues competition in the computer-use agent space, where Anthropic released a version of Claude with computer use last year.

Developers can access the model now in preview, with a demo environment hosted by Browserbase. The release sparked discussion among developers about production readiness, with some noting that current implementations can be slow and may be replaced by standard API calls.