Back to rankings

Gemini API: Computer Use Reaches General Availability

Gemini CLIView original changelog

Google moved the Computer Use tool out of preview and into general availability across both the Interactions API and the GenerateContent API. Developers can now build agents that look at screenshots and issue clicks, typing and scrolling actions in browser, desktop and mobile environments without relying on a preview-labelled feature. Computer Use includes built-in safety decisions and confirmation handling, so apps stay in control of risky actions.

Key Takeaways

  • Computer Use is out of preview, with Google declaring general availability on October 30, 2026 for a tool that launched as a Gemini 3.5 Flash preview in June.
  • The GA covers both the Interactions API and the GenerateContent API, so teams on the legacy interface are not forced to migrate to use it.
  • The tool runs as a screenshot-in, action-out loop, where the model returns a click or keystroke and the client executes it and sends back a new screenshot.
  • It supports browser, mobile and desktop environments, including actions such as drag and drop, long press, hotkeys and scrolling.
  • Seven built-in safety policy categories and a require_confirmation decision keep risky actions such as payments, account creation and legal agreements under human control.
  • Google still recommends sandboxing, allowlists and logging, and prompt injection detection is opt-in and needs Gemini 3.5 Flash-Lite or later.

Computer Use Is No Longer in Preview

Google announced on October 30, 2026 that the Computer Use tool in the Gemini API is now generally available. The release covers both the Interactions API and the GenerateContent API, so teams using either interface can rely on the tool without the preview label. Computer Use first arrived as a public preview built into Gemini 3.5 Flash in June 2026, after a standalone Gemini 2.5 Computer Use model launched in October 2025.

How Computer Use Works

Computer Use lets a Gemini model operate a user interface through a loop that the developer's application runs. The app sends the model a prompt and a screenshot. The model replies with a function call, such as a click or a keystroke, along with an intent that explains the action. The client executes that action, captures a new screenshot and returns it as a function result. The loop repeats until the task is finished.

The model returns normalized coordinates on a 0 to 999 grid, and the client code scales them to the actual viewport. Google's documentation uses Playwright as its example handler for browser control.

Supported Environments and Actions

According to the current documentation, the tool covers three environments:

  • Browser: click variants, mouse down and up, move, type, drag and drop, wait, key presses and hotkeys, screenshot, scroll, and navigate, back and forward.
  • Mobile: open app, click, list apps, wait, back, type, drag and drop, long press, key press and screenshot.
  • Desktop: largely the browser action set, without the navigation actions.

Developers can exclude predefined actions and add custom functions of their own, for example a human handoff function that pauses the agent and asks a person to take over.

Models

The current Computer Use documentation lists Gemini 4 Argon (gemini-4-argon), Gemini 3.8 Flash (gemini-3.8-flash, the recommended model), Gemini 3.5 Flash-Lite (gemini-3.5-flash-lite) and Gemini 3 Flash Preview (gemini-3-flash-preview) as supported models. The October 30 changelog entry itself is brief and does not list models, so the list comes from the documentation page rather than the announcement.

Safety and Confirmation Controls

Computer Use ships with seven built-in policy categories covering financial transactions, sensitive data modification, communications, account creation, data modification, consent management and legal agreements. Each function call can carry a safety decision of allowed, require_confirmation or blocked. When a call requires confirmation, the app must ask the user and then acknowledge the action with a safety_acknowledgement.

Developers can override policies with disabled_safety_policies, but the model may still return a require_confirmation decision, so applications need to handle it. Prompt injection detection is opt-in, defaults to off, and requires Gemini 3.5 Flash-Lite or later. Google warns that Computer Use may contain errors and security vulnerabilities and recommends sandboxing, allowlists, logging and human oversight.

What Changes for Developers

For anyone who held off on shipping browser or UI automation because the tool was labelled preview, general availability is the signal that the interface is stable enough for production planning. The documentation examples use the Interactions API, which Google describes as its default interface for new projects, while the GenerateContent API remains supported. The announcement does not mention new pricing or rate limits.