In the Gemini API changelog updated on 30 October 2026, Google announced that the Computer Use tool has reached general availability, covering both the Interactions API and the GenerateContent API. That means the ability to let a model look at screen captures and click and type directly has moved from preview to a production-ready tool. Computer Use first appeared as a public preview on 24 June 2026, already supporting browser, mobile, and desktop environments, along with intent explanations, configurable safety policies, and prompt injection detection.
Computer Use works as a loop. The developer first sends the model the user instruction, configuration, and a screenshot of the current screen, and the model returns an interface action after looking at the image, such as a click, a scroll, or a keystroke, together with an explanation of the intent behind that step and a safety decision. Safety decisions come in three grades: allowed, requiring user confirmation, or blocked outright. The developer's client program carries out the allowed action, then captures a new screenshot and sends it back to the model, repeating the cycle until the task is finished or stopped. The model itself only looks and decides; the actual work is done by the execution environment on the developer's side, and the official documentation requires running the agent inside a sandboxed virtual machine or container, with a Docker-based reference implementation offered as a starting point. Typical uses listed in the documentation include repetitive data entry and form filling, automated testing of web applications, and gathering product information, prices, and reviews across websites.
Generally available does not mean ready out of the box
General availability settles the question of interface stability, not the two hardest parts: execution and safety. Developers still have to build the client that handles screenshots, coordinate conversion, and action execution, deal with logins, captchas, and error pages, and decide which actions must first pass human confirmation. Prompt injection detection has to be turned on deliberately; it scans screenshots for hidden adversarial instructions, but the documentation does not promise to catch every injection. For users, the change will show up mainly on the supply side: more applications will claim to operate a computer on your behalf to book tickets, fill in forms, and compare prices, and the test will no longer be whether a model can read an interface, but whether it stops to ask before acting on your account, and whether its actions can be checked afterwards. This update gave no new pricing or benchmark results in the changelog, so before migrating, developers still need to measure success rates and per-task cost on their own tasks, rather than relying on the words generally available alone.
