Gemini API Computer Use is GA: gate every risky click

Gemini API Computer Use is now generally available. On October 30, Google moved the tool to GA across the Interactions and GenerateContent APIs. That makes it a supported way to build agents that drive a browser, a phone, or a desktop app by looking at the screen and clicking. For a small business, that's the path to automating the software that never got an API. It's also a model clicking buttons in your accounts, so the safety settings are the part to read first.

What actually happened

The Gemini API changelog entry for October 30, 2026 is one line: Computer Use is released to general availability across the Interactions and GenerateContent APIs. It took about a year to get here. Google launched a Gemini 2.5 Computer Use preview in October 2025, added it to Gemini 3 preview models in January, and in June shipped a Gemini 3.5 Flash preview with browser, mobile, and desktop environments, configurable safety policies, and prompt injection detection.

The Computer Use docs show how it works. You pass a computer_use tool with an environment (the example uses "browser" on gemini-3.8-flash), send a screenshot, and get back UI actions to run. Your code runs them and sends back the next screenshot. The Python examples need google-genai 2.7.0 or later.

Two controls matter most:

• safety_decision: each proposed action is classified as allowed, require_confirmation (needs a human yes), or blocked. On a confirmation, your app asks the user and returns an acknowledgement before the agent continues.
• Prompt injection detection: the docs examples turn it on with enable_prompt_injection_detection. Set it yourself; don't assume it's on.

Google doesn't list Computer Use pricing on the docs page, so meter usage before you scale it.

Why Gemini Computer Use matters for your business

Every business has one tool with no API: the supplier portal, the old POS back office, the insurance site you key claims into by hand. Computer-use agents are how that work gets automated. GA means a supported product you can build on, not a demo.

GA does not make it safe by default. Our rules for any screen-driving agent:

1. Wire up require_confirmation properly. Payments, sends, account changes, and accepting terms go to a human. Don't auto-acknowledge.
2. Give the agent its own account. Narrow permissions, nothing it doesn't need.
3. Turn on injection detection. A web page can carry instructions. Treat everything on screen as untrusted.
4. Log every action and screenshot. If you can't replay a run, you can't debug or audit it.

Sources: Gemini API changelog, Gemini API: Computer Use.