Gemini Computer Use goes GA: Google's AI can now click, type and scroll

Google has made Gemini Computer Use generally available (GA). The release lets developers build AI agents that operate browser, mobile, and desktop interfaces: the model “sees” the screen through screenshots, interprets what is on it, and suggests actions such as clicking, typing, scrolling, and dragging.

The tool is available through the Interactions and GenerateContent APIs, with no breaking changes from the preview — teams already testing it do not need to migrate anything. Google recommends Gemini 3.8 Flash for Computer Use; for workflows that need deeper reasoning, it points to Gemini 4 Argon. Gemini 3.5 Flash-Lite and Gemini 3 Flash Preview are also supported.

How it works

It is important to understand the tool’s limit: Gemini only proposes each action. Execution stays with the developer’s application, which uses tools such as Playwright to turn normalized coordinates (0 to 1000) into real clicks and send a fresh screenshot back to the model. The loop repeats until the task is done or hits a developer-defined limit.

Safety in every action

Each response includes an intent field explaining why the model chose that action, and a safety decision classifying the action as allowed, blocked, or requiring user confirmation. Built-in policies cover financial transactions, changes to sensitive records, autonomous messaging, account creation, and accepting terms. An optional prompt-injection detection scan of screenshots can flag malicious content trying to steer the agent.

What it can do

Suggested uses for Gemini Computer Use include automating repetitive data entry and form filling, automated testing of web applications, and research across websites — for example, comparing product prices and reviews across online stores before a purchase.

Gemini Computer Use carries no separate tool fee: billing follows each model’s standard token rates on the paid tier, with no free-tier access. Google provides a reference Docker environment on GitHub and a hosted Browserbase demo for anyone who wants to try the interaction loop before building their own runtime. The company recommends human supervision for consequential tasks, since detection can miss attacks and the model can pick the wrong action.

Image credit: diagram of the Computer Use interaction loop, from the official Google AI Studio documentation (Google).

Sources

Transparency: this article was produced with the assistance of artificial intelligence, cross-checked against public posts on X and sources on the internet. Please consult the original sources to confirm the details.

Muse