Gemini API Computer Use GA: an integration guide to client-side actions and user approval

What reached general availability?

Google's release notes announced Computer Use general availability on October 30, 2026, across the Interactions API and GenerateContent API. GA describes the tool's release status; it does not grant access to every Google AI Studio project, API key, region or plan, or establish API access through a consumer Gemini subscription. Before launch, check model access and supported regions for the project you will actually use.

The Computer Use guide lists gemini-4-argon, gemini-3.8-flash, gemini-3.5-flash-lite and gemini-3-flash-preview as supported models. Tool GA and a model's release stage are separate: gemini-3-flash-preview remains in Preview.

Choose a model and execution environment

Google recommends gemini-3.8-flash for Computer Use. gemini-4-argon is positioned for complex reasoning and agent workflows, while gemini-3.5-flash-lite supports lower-latency, lower-cost tasks; gemini-3-flash-preview is still a preview model. Compare accuracy, elapsed time and usage on your own task samples. The recommendation is not a performance test of your application.

The tool can be configured for browser, mobile or desktop environments. The mobile environment is optimized for Android. You provide a controlled runtime, a source of screenshots and a client that can execute actions in that environment; the API does not connect to or operate your local desktop by itself.

The model suggests actions; your client controls execution

After you send a request with the current screen state, the model returns a Computer Use function call. It can include actions such as click, type or scroll, an intent explaining the suggested action, and a safety_decision. Validate the function name, argument types and target scope before passing it to your browser, mobile or desktop executor.

Intent explains the model's suggestion; it is not an authorization token. Your application still decides whether an action fits the task, destination allowlist and user permissions. Do not treat model output as a command that can write directly to business systems.

Connect screenshots and function_result in a loop

Execute an accepted action in the selected environment, capture a fresh screenshot and return the execution result as function_result. The model can then propose its next step from the updated screen. Keep both model responses and function results in the conversation history, and stop when the task is complete, a step limit is reached or the request times out.

Coordinates use a normalized 0–999 range. Convert them to pixels using the viewport width and height that correspond to the current screenshot, and make sure the screenshot and execution window still have the same dimensions and state. If a dialog appears, the page changes or an action is unfamiliar, pause and reassess instead of repeating a blind click.

For a hypothetical task on your own test site, the model might suggest opening a menu to read a field. The client checks that the destination is still an authorized page, clicks, captures a screenshot and returns the result. This is an integration example, not a test performed on a live site. If a turn returns multiple function calls, provide a corresponding result for each call.

Treat safety_decision as an execution gate

When safety_decision is allowed, still apply your application's permission and destination checks. Stop an action marked blocked. For require_confirmation, pause and ask the end user. Only after the user actually confirms should you include safety_acknowledgement in the corresponding function_result. A configured safety override does not guarantee that a later response will not require confirmation.

Gemini 3.5 Flash-Lite and later models support optional screenshot prompt-injection detection, which is off by default. Enabling it adds a check but does not replace human approval, a strict action allowlist or an isolated runtime. Start by validating the full loop on low-risk tasks that you own or are explicitly authorized to perform.

Estimate token and runtime costs

The official pricing table lists Computer Use as Not available on the free tier; paid usage is charged at the selected model's regular token rates. There is no fixed price per click. For Standard gemini-3.8-flash, input is $0.75 per million tokens and output is $3.75, including thinking tokens, through December 31, 2026. Starting January 1, 2027, those rates are $1.50 and $7.50. Rates can change, so check the pricing page again before launch.

Screenshots, repeated conversation history and multi-turn reasoning affect token usage. A browser, container, virtual machine or mobile device can add separate runtime costs. Check the supported-region list for the project you will use; tool GA does not expand Google's supported regions. Before sending screenshots that may contain sensitive information, review request logging and storage settings and read Gemini API request logging and storage (https://puppyip.com/en/resources/gemini-api-request-logging-storage-guide).