Google has taken one of the Gemini API's agent features out of preview. An October 30 entry in the Gemini API release notes states that the Computer Use tool is now generally available in both the Interactions API and the GenerateContent API, a status that signals Google considers it ready for production applications rather than experiments.
Computer Use is the building block for agents that operate software through its interface instead of through APIs. The developer sends the model a task, a screenshot of the current screen and the target environment, and the model replies with a proposed action such as a click, a scroll or text to type. Google does not run those actions. The developer's own code carries them out, for example with the Playwright browser automation library, converts the model's normalized coordinates on a 1000 by 1000 grid into real pixels, captures a fresh screenshot and repeats the loop until the task is finished.
The road to this release was long. Google first offered a Gemini 2.5 Computer Use preview in October 2025, added the tool to gemini-3-pro-preview and gemini-3-flash-preview in January 2026, and in June 2026 launched a public preview in Gemini 3.5 Flash that introduced support for browser, mobile and desktop environments, intent explanations and the current safety controls. The documentation now lists four models for the tool: Gemini 4 Argon for the hardest reasoning, Gemini 3.8 Flash as the recommended default, the cheaper Gemini 3.5 Flash-Lite and Gemini 3 Flash Preview.
Much of the documentation is about keeping such agents in check. With Gemini 4 and Gemini 3.x models, each proposed action carries an intent field explaining why the model chose it, and an internal safety system may mark the step as allowed, as requiring the user's approval or as blocked. Built-in policy categories cover payments and checkout, changes to health, financial or government records, sending emails or chat messages, creating accounts, deleting or sharing data, cookie and privacy prompts, and accepting terms of service. Developers can switch some of these off, but Google warns that the model may still ask for confirmation. From Gemini 3.5 Flash-Lite onward, an optional check scans each screenshot for hidden instructions planted to hijack the agent and stops execution when it finds them; it is off unless developers turn it on.
Google is also explicit about the remaining risk. The documentation warns that the tool may contain errors and security vulnerabilities, recommends close supervision for important tasks and advises against using it for critical decisions, sensitive data or actions whose mistakes cannot be undone. It suggests running agents inside a sandboxed virtual machine or container and starting each task from a clean state, and developers can remove built-in actions and register their own functions when a human needs to stay in the loop.
The release so far consists of a short changelog line and an updated guide; no blog post, pricing change or benchmark results accompanied it on the day. For teams building browser or desktop agents, the practical change is that a capability Google kept in preview for about a year is now part of its supported API surface, which puts it in direct competition with the computer-use tools offered by Anthropic and OpenAI.