Introducing the Gemini 2.5 Computer Use model

Google DeepMind just shipped something it's been teasing since earlier this year: a version of Gemini 2.5 Pro trained specifically to operate graphical interfaces. Not APIs, not structured data feeds — actual buttons, dropdowns, forms, and login screens, the messy stuff humans deal with every day. The model is called Gemini 2.5 Computer Use, it's in preview now, and developers can grab it through the Gemini API in Google AI Studio or Vertex AI.

The pitch is straightforward. Plenty of useful software still has no clean API, so an agent that wants to book an appointment, fill out a form, or drag sticky notes into the right columns on a kanban board has to see the screen and act on it, one click at a time. Google's demos lean into this: one shows an agent pulling pet records from a signup page and scheduling a spa visit around a specific vet's availability, another has it tidying a chaotic virtual sticky-note board into categories a user defined. Neither task involves an API call. Both involve reading pixels and making decisions.

Under the hood, the system runs a loop. The model takes in a screenshot, the current URL, a history of recent actions, and the user's request, then spits out a function call — click here, type this, scroll there. That action gets executed, a fresh screenshot comes back, and the cycle repeats until the task finishes, something breaks, or a safety check intervenes. Google says the model currently outperforms competing tools on Online-Mind2Web and other web and mobile control benchmarks while running with noticeably lower latency, based on both its own testing and third-party evaluation from Browserbase. It's tuned mainly for browsers right now; mobile UI control works well too, but desktop OS-level control isn't there yet.

Safety is the part Google spends real time on, and for good reason — an agent that can click anything on a screen can also click the wrong thing, whether that's approving a purchase, bypassing a CAPTCHA, or getting fooled by a prompt injection buried in a webpage. Google's response is a layered one: safety training baked into the model itself, plus an external per-step safety service that reviews each proposed action before it fires, plus developer-set system instructions that force confirmation or outright refusal on higher-stakes moves. Internally, earlier versions of this model already power Project Mariner, the Firebase Testing Agent, and some of the agentic features inside AI Mode in Search, so this isn't a lab experiment dressed up for a blog post — it's already doing production work.