Playbooks · Coding tasks
Gemini Computer Use is generally available
Written November 7, 2026
The Gemini API's Computer Use is GA: point a model at a screen and let it click, type, and scroll. What it exposes, what it costs, and where it needs a human.
What went GA
The Gemini API release notes for October 30, 2026 carry a one-line entry: Computer Use is generally available across the Interactions and GenerateContent APIs. That is the whole announcement — the entry names no model, no price, no region, and no limit. What changed is the status, not the mechanics: the tool leaves preview.
What the tool does
Computer Use is a model tool for browser, mobile, and desktop environments. Using screenshots, the model can see a computer screen and act by generating specific interface actions such as mouse clicks and keyboard inputs. The docs' example enables the computer_use tool with an environment of browser and the model gemini-3.8-flash.
The docs list four models for Computer Use: Gemini 4 Argon, Gemini 3.8 Flash (the recommended model), Gemini 3.5 Flash-Lite, and Gemini 3 Flash Preview. The 3.8 Flash model is the one in the docs' example, and the recommended choice for high-accuracy interface interaction and reliable tool calling.
The loop runs in four steps, and the division of labour is the point of the design:
- Your application sends a request with the Computer Use tool, its configuration, the prompt, and a screenshot of the current screen.
- The model reads the screen and returns a suggested function call — a click, scroll, or keystroke. For Gemini 4 and Gemini 3.x models, the response also includes an
intentfield explaining why the model chose that action. - Your client-side code parses the call, scales the coordinates to the viewport, and executes the action with automation tools such as Playwright.
- Your application captures a new screenshot and sends it back for the next step.
The model suggests; your application executes. GA does not move the execution to Google's side, and it does not remove the application layer between a model proposal and a changed screen.
Where a human still belongs
Every response can carry a safety_decision from an internal safety system that classifies the action as allowed, require_confirmation, or blocked. If an action is allowed — or the user confirms it — your client runs it; if it is blocked, your client halts or handles the interruption. Prompt-injection detection is opt-in screenshot scanning for hidden adversarial instructions, disabled unless you enable it.
Google recommends running the agent in a sandboxed virtual machine or container to isolate it from the host system, and the reference implementation includes a Docker-based sandbox as a starting point. The docs also carry a plain warning: Computer Use may contain errors and security vulnerabilities, so supervise closely for important tasks and avoid using it for critical decisions, sensitive data, or actions where serious errors cannot be corrected.
What it costs
Computer Use is not available on the Free tier. On the Paid tier it is charged as regular tokens per model pricing — there is no separate Computer Use fee, so the cost follows the model you pick. The Gemini Developer API pricing page lists the tool that way, using standard Gemini 3.8 Flash pricing as its example.
Limits
- Nothing here reports a Computer Use session actually run. The GA status, the action loop, and the pricing above come from Google's changelog, tool docs, and pricing page.
- GA establishes availability, not performance. Google's own docs warn the tool can make mistakes, and no accuracy or reliability figures come with the release.
Related playbooks
- Copilot Fusion is generally available: what it does to your AI-credit bill
- Tencent Hy4: vision and computer use, weights pending
- Gemini Decision: a model and API for choices
- Copilot computer use: drive the GUI
Sources
- Release notes — Gemini API docs — Google
- Computer use — Gemini API docs — Google
- Gemini Developer API pricing — Google
Back to: Coding playbooks hub · Also: Agent-desktop drives apps via the access tree · Tencent Hy4: vision and computer use, weights pending · Gemini API now takes Office files and returns files