# Google Makes the Gemini API's Computer Use Tool Generally Available

Times of AI Desk · 2026-11-01 · Products

[https://timesof.ai/2026/11/gemini-api-computer-use-generally-available](https://timesof.ai/2026/11/gemini-api-computer-use-generally-available)

> Google moved the Gemini API's Computer Use tool to general availability on October 30 across the Interactions and GenerateContent APIs. Its documentation lists Gemini 4 Argon, Gemini 3.8 Flash, Gemini 3.5 Flash-Lite and Gemini 3 Flash Preview as supported, with browser, mobile and desktop environments and opt-in prompt-injection scanning of screenshots.

General availability matters less for what the tool can do than for who can depend on it. Developers building agents that drive screens on Gemini can now treat Computer Use as a supported product rather than a preview, roughly ten weeks after Anthropic made the same move with [computer use on the Claude Platform](https://timesof.ai/2026/08/anthropic-claude-platform-computer-use-ga). Google's own documentation still warns against using it for critical decisions.

## What changed

"Released the Computer Use tool to general availability (GA) across the Interactions and GenerateContent APIs," Google wrote in the Gemini API changelog entry dated October 30, 2026. The tool lets developers build browser, mobile and desktop agents: the model looks at screenshots and returns UI actions such as clicks and keystrokes, which the developer's own code carries out, according to the documentation.

The documentation lists four supported models: Gemini 4 Argon, Gemini 3.8 Flash, which Google calls the recommended model for computer use, Gemini 3.5 Flash-Lite and Gemini 3 Flash Preview. For Gemini 4 and Gemini 3.x models, Google lists multi-environment support, actions that carry an "intent" field explaining each step, configurable safety policies and opt-in screenshot scanning to detect hidden prompt-injection instructions. [Gemini 4 Argon](https://timesof.ai/2026/09/google-gemini-4-argon) launched in September with access initially limited to trusted cyber defenders.

## The built-in safety categories

The documentation lists seven safety policy categories for Gemini 4 and Gemini 3.x models that decide when the model asks for user confirmation before acting. They cover financial transactions such as payments and checkout, changes to health, financial or government records, sending emails or chat messages, creating accounts, file-system changes and deletions, cookie and privacy prompts, and accepting terms of service or other contracts. When one applies, the model returns a safety decision of "require_confirmation," and the developer's code is expected to ask the end user and pass back an acknowledgement. Developers can override "select policies," but Google says overrides "indicate your preference" and the model "may still return a safety_decision with require_confirmation in some cases," so applications should always handle it. The sample system prompt in the documentation also tells the agent never to solve or bypass CAPTCHAs.

## Analysis: the confirmation step stays in Google's hands

Our reading: by putting these categories in the model's own output, Google keeps a say over the riskiest actions even after a developer configures the tool. A developer can loosen the defaults but cannot fully switch them off, because the model may still ask for confirmation. That makes the tool easier to defend when an agent buys, sends or signs something by mistake. It is a softer approach than OpenAI's [Lockdown Mode](https://timesof.ai/2026/06/openai-lockdown-mode-prompt-injection-protection), which switches outbound agent tools off entirely, but it rests on the same idea: some actions need a person. It also means fully unattended flows that pay, message or sign up will hit confirmation steps the developer does not fully control. Teams moving from a preview build should plan for a human in the loop on those actions rather than treat the prompts as rare.

## Limits

- Google's changelog and documentation are the only substantive sources. As of November 2 we found no Google blog post and only brief aggregator write-ups that restate the changelog. The pages we read do not give pricing or rate limits specific to the tool.
- The documentation warns that Computer Use "may contain errors and security vulnerabilities" and recommends close supervision and avoiding tasks involving critical decisions, sensitive data or errors that cannot be corrected.
- We have not tested the tool, and Google publishes no accuracy figures for it on these pages.

## Sources

- [Google AI for Developers: Gemini API release notes (October 30, 2026)](https://ai.google.dev/gemini-api/docs/changelog)
- [Google AI for Developers: Computer use (last updated October 30, 2026)](https://ai.google.dev/gemini-api/docs/computer-use)