Last updated: November 3, 2026
Gemini Computer Use became generally available in the Gemini API on October 30, 2026. It lets Gemini 4 Argon, Gemini 3.8 Flash and other supported models read screenshots and return clicks, typing and other UI actions for browser, mobile and desktop agents. Your own code runs each action.
Contents
- What changed at a glance
- How the agent loop works
- How to set up a browser agent
- Actions by environment
- Safety policies and confirmations
- Pricing and models
- GA tool compared with earlier versions
- Limitations
- Frequently asked questions
Google announced the change in the Gemini API release notes: the Computer Use tool moved to general availability across the Interactions and GenerateContent APIs.
What changed at a glance
| Item | Detail |
|---|---|
| Status | Generally available since October 30, 2026 |
| APIs | Interactions API and GenerateContent API |
| Recommended model | Gemini 3.8 Flash (gemini-3.8-flash) |
| Environments | Browser, mobile (Android-optimized) and desktop |
| New in responses | An intent field that explains each action (Gemini 4 and 3.x models) |
| Safety | Seven built-in policy categories, overrides and opt-in prompt injection detection |
| Billing | Regular tokens at the model’s rates; not available on the free tier |
For background on the recommended model, see our explainer on what Gemini 3.8 Flash is. If you want the most capable option, our Gemini 4 Argon guide covers access and limits.
How the agent loop works
According to the Computer Use documentation, the tool works much like function calling. The model sees the screen through screenshots and suggests an action, but it never touches your machine directly. You build the client-side environment that carries out each step.
Each turn has four parts. Your app sends the prompt, the tool configuration and a screenshot. The model replies with a function call, such as a click with x and y coordinates, plus an intent explaining why. Your code runs the action, captures a new screenshot and returns it in a function result. The loop repeats until the task ends.
Coordinates come back on a normalized 0 to 999 grid. You do not pass the display size in the request; your code scales the values to your real viewport before clicking.
How to set up a browser agent
The documentation uses Python, the google-genai SDK (version 2.7.0 or higher) and Playwright. A minimal setup looks like this:
- Run the agent in a sandboxed VM or container. Google’s reference implementation includes a Docker-based sandbox you can start from.
- Install the packages with
pip install google-genai playwright, then runplaywright install chromium. - Start a Playwright browser with a fixed viewport, such as 1440 by 900 pixels, and open a starting page.
- Call
client.interactions.createwithmodel="gemini-3.8-flash"andtools=[{"type": "computer_use", "environment": "browser"}]. Add"enable_prompt_injection_detection": Trueif you want screenshot scanning. - Parse each returned function call, scale its coordinates and run it in Playwright.
- Take a screenshot and send it back in a function result. Cap the number of turns, as the sample loop does with a limit of five.
The code is on GitHub in Google’s computer-use-preview reference implementation. If you prefer the terminal, our Gemini API CLI guide shows another way to call the Interactions API.
Actions by environment
Gemini 4 and Gemini 3.x models support three environments, set with the environment value in the tool configuration. The action sets differ.
| Environment | Core actions | Environment-specific actions |
|---|---|---|
| Browser | click, double_click, triple_click, middle_click, right_click, mouse_down, mouse_up, move, type, drag_and_drop, wait, press_key, key_down, key_up, hotkey, take_screenshot, scroll | navigate, go_back, go_forward |
| Mobile (Android-optimized) | click, type, drag_and_drop, wait, press_key, take_screenshot, go_back | open_app, list_apps, long_press |
| Desktop | The same cursor and keyboard actions as the browser, including scroll and hotkey | No URL navigation actions |
You can also add your own functions. Google’s example excludes the built-in click action with excluded_predefined_functions and registers a custom yield_to_user function, so the agent hands control back to a person when a step is unsafe or unclear, such as a two-factor code.
Thinking levels also apply. Google says lower thinking levels generally give a good balance of speed and action quality for standard automation tasks.
Safety policies and confirmations
Gemini 4 and 3.x models include built-in safety categories that decide when a person must confirm an action:
- FINANCIAL_TRANSACTIONS: payments, checkout and regulated goods.
- SENSITIVE_DATA_MODIFICATION: health, financial or government records.
- COMMUNICATION_TOOL: sending emails, chat messages or drafts.
- ACCOUNT_CREATION: registering new accounts.
- DATA_MODIFICATION: file system changes, data sharing and deletion.
- USER_CONSENT_MANAGEMENT: cookie banners and privacy prompts.
- LEGAL_TERMS_AND_AGREEMENTS: accepting terms of service or contracts.
You can relax a category with disabled_safety_policies, for example data_modification for a log cleanup agent. Google notes that the model may still ask for confirmation, so your app must always handle safety decisions.
When a response includes a safety_decision of require_confirmation, prompt the user. If they approve, set safety_acknowledgement in the function result. If an action is blocked, stop or handle the interruption.
Pricing and models
The Gemini API pricing page says Computer Use is not available on the free tier. On the paid tier it is charged as regular tokens at the chosen model’s rates, with no separate tool fee listed. For broader rate details, see our Gemini API pricing breakdown.
| Model | Model string | Notes from Google |
|---|---|---|
| Gemini 4 Argon | gemini-4-argon | Most capable; complex reasoning and agentic UI workflows |
| Gemini 3.8 Flash | gemini-3.8-flash | Recommended for computer use; $0.75 input and $3.75 output per 1M tokens through December 31, 2026, then $1.50 and $7.50 |
| Gemini 3.5 Flash-Lite | gemini-3.5-flash-lite | Low latency and lower cost |
| Gemini 3 Flash Preview | gemini-3-flash-preview | Preview model |
GA tool compared with earlier versions
Computer use started as a standalone Gemini 2.5 computer use model. On June 24, 2026, Google made it a built-in tool in Gemini 3.5 Flash, added adversarial training against prompt injection and introduced optional confirmation and injection-stop safeguards.
| Aspect | June 2026 (3.5 Flash) | October 2026 GA |
|---|---|---|
| Packaging | Built-in tool in Gemini 3.5 Flash | Built-in tool across four listed models |
| APIs | Gemini API and Gemini Enterprise Agent Platform | Interactions and GenerateContent APIs, now GA |
| Environments | Browser, mobile and desktop | Browser, mobile and desktop, with documented action sets |
| Safety controls | Optional confirmation and injection-stop safeguards | Seven policy categories, overrides and opt-in injection detection |
If you are planning a migration, note that Google is retiring older Gemini 2.5 models. The release notes say requests to gemini-2.5-pro and gemini-2.5-flash will route to newer models from December 1, 2026. Our Gemini usage limits guide covers app-side caps.
Limitations
- You run the actions. The API returns suggestions only; reliability depends on your executor and screenshots.
- Errors are possible. Google warns the tool may contain errors and security vulnerabilities and advises close supervision for important tasks.
- Avoid critical work. Google advises against tasks with critical decisions, sensitive data or errors that cannot be corrected.
- Keep the screen stable. Unexpected pop-ups and layout changes can confuse the model, so start each task from a known, clean state.
- Log everything. Google recommends logging screenshots, suggested actions, safety responses and executed actions.
Frequently asked questions
Is Gemini Computer Use generally available?
Yes. Google’s Gemini API release notes say the Computer Use tool reached general availability on October 30, 2026, across both the Interactions API and the GenerateContent API.
Which models support Gemini Computer Use?
The documentation lists four: Gemini 4 Argon (gemini-4-argon), Gemini 3.8 Flash (gemini-3.8-flash), which Google recommends for computer use, Gemini 3.5 Flash-Lite (gemini-3.5-flash-lite), and Gemini 3 Flash Preview (gemini-3-flash-preview).
How much does Gemini Computer Use cost?
Google’s pricing page says Computer Use is not available on the free tier. On the paid tier it is charged as regular tokens at the selected model’s rates. For Gemini 3.8 Flash that is $0.75 per million input tokens and $3.75 per million output tokens through December 31, 2026.
Does Gemini Computer Use click and type for me?
No. The model returns suggested actions such as click, type or scroll as function calls. Your own client code, for example a Playwright script, must execute each action, capture a new screenshot and send it back.
How do I turn on prompt injection detection?
Add enable_prompt_injection_detection set to true in the computer_use tool configuration. It is off by default and works with Gemini 3.5 Flash-Lite or later. When it finds hidden adversarial instructions in a screenshot, it blocks execution.
Sources
- Gemini API release notes (Google AI for Developers)
- Computer use documentation (Google AI for Developers)
- Gemini Developer API pricing (Google AI for Developers)
- Introducing computer use in Gemini 3.5 Flash (Google blog, June 24, 2026)
- google/computer-use-preview reference implementation (GitHub)
Image: Google Gemini logo by Google LLC, public domain (trademarked), via Wikimedia Commons. Gemini is a trademark of Google LLC. FreeAnswers is not affiliated with or endorsed by Google.
