Google DeepMind ships a computer-use model via the Gemini API

Google DeepMind released the Gemini 2.5 Computer Use model, a specialised version of Gemini 2.5 Pro built to let developers create agents that interact directly with graphical interfaces — clicking buttons, typing into fields, scrolling and navigating dropdowns — rather than acting only through APIs designed for machine use. The model operates in a loop, receiving a screenshot of the current screen state and returning the next UI action to take, mimicking how a person would operate a browser.

DeepMind said the model could fill out forms, operate behind login screens and manipulate interactive page elements, and built in a checkpoint for high-stakes actions such as making a purchase, requiring explicit user confirmation before proceeding. The company reported that it led rival models on browser-control benchmarks including Online-Mind2Web, WebVoyager and AndroidWorld, combining higher task-completion rates with lower latency, though it did not publish independently verified comparative figures. DeepMind was explicit that the model was optimised for web browsers, showed promise but less optimisation for mobile interfaces, and was not tuned for controlling a desktop operating system directly.

The release made Google DeepMind, OpenAI and Anthropic’s frontier agent offerings — each of which had shipped comparable browser- or computer-use capabilities within the same year — converge on the same underlying idea: giving a model visual access to a screen and a small action space, rather than requiring every target application to expose a bespoke API, as the general-purpose route to agentic software control. It also extended the security and reliability concerns that had accompanied earlier computer-use releases, since a model acting through screenshots and clicks is harder to constrain than one restricted to a defined set of API calls.