Gemini Computer Use Is Generally Available: Models, Costs and Safety Controls for Browser Agents

Gemini computer use became generally available on October 30, 2026, according to Google’s Gemini API changelog. The tool lets a Gemini model look at screenshots and return clicks, typing and scrolling for your code to carry out. It now works on four models across browser, desktop and mobile environments.

For agencies that build browser automation, GA means the tool is out of preview. This post covers what GA includes, which models support it, what a run costs, and the safety controls Google documents. It also covers the gemini-api command-line tool that launched in beta three days earlier.

The short version: Gemini computer use is billed at the normal token prices of the model you pick, and every step sends a fresh screenshot. Gemini 3.8 Flash is the low-cost choice today, but its price doubles on January 1, 2027. Google’s own safety model depends on your code: you must handle its confirmation requests, run the agent in a sandbox, and decide which actions need a person. Read the safety section before you point it at a client’s live site.

What GA means for Gemini computer use

Google’s Gemini API changelog lists the October 30 entry as "Computer Use generally available (GA)." It says the tool is GA on both the Interactions API and the GenerateContent API.

The Computer Use documentation describes how it works. The model reads a screenshot and returns UI actions. Your client code runs them. Google compares it to function calling: you build and host the environment where the clicks happen.

The loop has four steps, per the docs:

1. Send the prompt, the tool setting and a screenshot.
2. Get back actions with screen coordinates and a short intent string.
3. Check any safety decision, then run the action.
4. Take a new screenshot, send it back, and repeat until the task is done.

Every action carries an intent, a plain description of what the model is trying to do. Log it. It is the closest thing to an audit trail the tool gives you.

GA also closes out the old preview model. Google’s deprecations page lists gemini-2.5-computer-use-preview-10-2025 with a shutdown date of July 28, 2026, and names Gemini 3.8 Flash as the replacement. Computer use is now a tool on general models, not a separate model.

Supported models and environments

The docs list four models under "Model versions": Gemini 4 Argon, Gemini 3.8 Flash, Gemini 3.5 Flash-Lite and Gemini 3 Flash Preview. The last one is labeled a "Preview model supporting computer use."

You declare the tool in an Interactions API call. The docs show a short form:

tools=[{"type": "computer_use", "environment": "browser"}]

The environment can be browser, desktop or mobile. Browser and desktop share most actions: clicks, typing, scrolling, drag and drop, key presses and screenshots. Browser adds navigate, go_back and go_forward. Mobile has its own set, including open_app and long_press. Coordinates come back on a normalized 1000 by 1000 grid that you scale to your screen.

For most agency work, browser is the one that matters. Typical jobs are checking forms after a CMS update, walking a donation flow, or moving content between two admin screens that have no API. Desktop control raises the stakes, since the agent can reach anything on that machine.

We covered which Gemini model fits which job in our October Gemini model guide. The Gemini 4 Argon GA post covers Argon’s limits.

What Gemini computer use costs

Google’s pricing page has no separate charge for computer use. We found only a link to the tool under Tools. You pay the model’s token rates, and each screenshot is image input.

Paid-tier standard prices per million tokens, as of November 2026:

Output prices include thinking tokens. The pricing page lists a free tier for 3.8 Flash and 3.5 Flash-Lite. Argon has no free tier.

Here is the math for an example task. Say each step sends 3,000 input tokens and returns 300 output tokens. On 3.8 Flash, that is about $0.0034 per step. A 40-step task costs about $0.14. On Argon, the same task costs about $0.36 today and about $0.72 after December 1.

Real runs can cost more. Screenshots add up, retries add steps, and long tasks carry more history. Run a test task, read the token counts in the response, and price from those. Our post on what Gemini 3.8 Flash costs per task explains why the per-task number matters more than the price per token.

Gemini computer use safety controls to turn on

Google gives you controls for Gemini computer use, and most of them work only if your code honors them.

Safety decisions. The model can return a safety_decision. The docs say: "If the safety_decision is require_confirmation, prompt the end user." If the person confirms, your code sets safety_acknowledgement in the result. If they decline, stop the run. A decision of blocked means your client should halt.

Built-in policy categories. The docs list categories such as financial transactions, sensitive data changes, communication tools, account creation, data changes, user consent and legal terms. You can turn some off with disabled_safety_policies. The docs warn the model may still ask for confirmation, and say: "Your application should always implement safety decision handling regardless of configured overrides."

Prompt injection detection. An optional flag, enable_prompt_injection_detection, is off by default. The docs say it applies to Gemini 3.5 Flash-Lite or later. Turn it on for any agent that reads pages you do not control.

Removing actions. excluded_predefined_functions takes built-in actions away. You can add your own functions, such as a yield_to_user step that hands control to a person.

Your own rules. Google’s sample system instruction tells the agent to get approval before purchases, sending messages, logins, accepting consent prompts and solving CAPTCHAs. That is a prompt you write, not a built-in guarantee.

Rules for Gemini computer use on client sites

For client work, three rules follow from the docs. Run the agent in a container with a test account, never a staff login. Require a person for anything that submits a form, spends money or sends a message. Keep the agent off production admin screens until you have watched it run on staging.

Our post on agentic browser security covers the wider risks of letting an AI agent act inside a browser.

Screenshots, logs and client data

Every step sends a screenshot, and screenshots show whatever is on screen. That can include donor names, member records or a student roster.

Check logging before you run Gemini computer use on real data. On October 26, Google turned request storage on by default for new paid-tier projects, matching the Interactions API, per the changelog. You can opt out per request with store=false or in project settings. Our post on the Gemini API logging default explains what is kept and for how long.

Sites are also getting stricter about bots. Our post on AI shopping agents being blocked covers donation and checkout pages that now turn agents away. An automation that works today on a client’s form may hit a bot check next month.

The gemini-api CLI beta

On October 27, Google launched the Gemini API CLI in beta. Google calls gemini-api "a stateless, zero-dependency command-line client built for AI agents with shell access." It works with the Interactions API to generate text and media, analyze files and run Gemini managed agents.

The page says it is "not an interactive coding agent like @google/gemini-cli." It is meant to run inside one, such as Claude Code, Codex or Cursor. It reads the GEMINI_API_KEY environment variable. The CLI page does not mention computer use.

Two cautions. The page warns that breaking changes can occur between minor versions, so pin a version in CI. The quick install pipes a script from GitHub into your shell. Read the script first, or install with Go, which the page also documents. Our July post on the move from Gemini CLI to Antigravity CLI covers Google’s other command-line tools.

How Anthropic and OpenAI compare

Anthropic offers computer use as computer_toolset_20260801, which needs no beta header on the Claude API, according to its computer use tool docs. Those docs say declaring it adds about 4,500 input tokens to a request, and that each screenshot costs roughly 1,000 to 1,800 input tokens. OpenAI’s computer use guide describes a computer tool in the Responses API, where "Screenshots count as image input." All three ask you to keep a person in the loop for consequential steps. We covered Anthropic’s desktop version in Claude computer use while you’re away, OpenAI’s browser agent in the ChatGPT Chrome extension post, and OpenAI’s lower-cost model in GPT-6.1 Terra.

Frequently Asked Questions