#gemini#computer use#google#agentes de ia#automatizacion#playwright

    Gemini Computer Use is now official: the AI that uses your browser and what it costs

    Gemini Computer Use is now generally available: the AI that clicks around your browser for about 10 cents a task. Rules and use cases.

    Gemini Computer Use is Google's tool that lets a Gemini model drive a browser by looking at screenshots, and on October 30, 2026 Google marked it generally available, in both of its APIs (Interactions and GenerateContent). By my own math, a 12-step task costs about 10 cents. Here is how it works, three uses, the math, and the five rules before you let it touch an account.

    What changed with Gemini Computer Use

    The headline is misleading: Gemini learning to use a computer has existed as a preview since January, and since June in Gemini 3.5 Flash. The news is that it's no longer a preview, the kind that can change without warning, and a business can put it into a real process. It matters today because newer models like Gemini 4 Argon and Gemini 3.8 Flash already run it. The documentation doesn't spell out what else changed on October 30: it only marks the move to general availability.

    A laptop with a miniature shopping cart

    How it works inside

    Gemini never touches your computer. The model only looks at a screenshot and answers with a written action ("click at this spot," "type this text"). The one who clicks is your code, for example with Playwright. That means you control where it runs, which accounts it sees, and when it stops. Coordinates come back normalized from 0 to 999 and your program scales them to the real screen. When someone says an AI got into your account, ask who wrote the code that made the click.

    How to use Gemini Computer Use step by step

    1. Open a browser controlled by your code and take a screenshot.
    2. Send it to the API with the goal in plain text (for example, "log into the supplier portal and tell me the status of order 42") and declare the Computer Use tool.
    3. The model returns an action, say a click on the search button.
    4. Your code runs it, waits for the page to load, and takes another screenshot.
    5. Send that back and the loop repeats until the model says it's done.

    It jams the first time if you capture before the page loads: wait for it to settle. And if the action touches something sensitive (a payment, accepting a contract), the model returns a confirmation request: ask a person and continue only with their yes.

    The math: how much a task costs

    According to Asana's 2022 index, 58% of the workday goes to work about work (hunting for information, switching apps, checking status). It doesn't measure portals, but it gives you the scale.

    My own calculation: a full-page screenshot is 6 tiles of 258 tokens, or 1,548. A 12-step task resends the history every time and adds up to about 120,000 tokens. At $0.75 per million with Gemini 3.8 Flash that's $0.09, and with the response, about $0.10. Twenty tasks a day for 22 days comes to about $42 a month. On January 1 that rate doubles, and reasoning counts as output: measure yours.

    A person shopping online on a laptop

    Three business uses (my own analysis)

    An online store with a supplier that has no API. Every morning the AI logs into the portal, looks up the ten best-selling products, copies price and stock, and drops them into a spreadsheet. The owner starts the day with current prices without opening ten tabs, and you can charge for it as a monthly maintenance plan.

    An agency maintaining client websites. Every morning it opens each site, fills in the contact form with test data, and checks that the success message shows up; if it fails, it sends a screenshot to whoever's responsible. It doesn't need to know the code, only what a visitor sees.

    A shipper checking several carriers' portals. The AI logs in, looks up the tracking number, reads the status, and builds a single dashboard. It's the hardest to copy because every portal asks you to log in: start with a read-only account, never one that can move money.

    Who it's not for

    • Anyone who already has an API: use it, it'll be faster, cheaper, and won't depend on guessing a button.
    • Anyone who needs instant answers: the loop is step by step and every step waits on the model.
    • Anyone who wants to turn it loose with money or contracts, unsupervised.

    The objection and what they don't tell you

    "Why pay for an AI when a Playwright script does the same thing for free?" A well-built script is cheaper, faster, and always does the same thing, but it breaks when the portal moves a button. My own rule: same task thousands of times on a stable site, write the script; different sites, frequent changes, or only a few times, the AI wins; if you're unsure, start with the AI and freeze whatever works.

    On "it's generally available now": what Google marked generally available is the tool inside the API, not that the model will get it right on your portal. The documentation itself warns the tool may contain errors and security vulnerabilities, and no published success rate was found. Until there's an independent test, run it yourself: 50 real tasks, five runs each. And prompt injection detection ships turned off by default.

    What breaks in practice

    • Slow pages or pages that change midway through the loop: the model acts on a stale screenshot.
    • Cookie banners, pop-ups, and confirmation requests, which the tool treats as sensitive cases.
    • Cost running away: a task that takes 20 turns costs far more than one that takes 12. Set a step cap.
    • Clicks in the wrong place from scaling the coordinates wrong.

    How to apply it in your business?

    The five rules: 1) an isolated browser, with a test account, no saved card, and no admin permissions; 2) an allowlist of sites; 3) a cap on steps and spend per task (say 15 steps); 4) human confirmation for anything that moves money, deletes something, or sends a message; 5) save every screenshot and every action. Turn on prompt injection detection and start with a read-only flow, like checking prices. To see another AI that drives your screen, read the full breakdown of GPT-6.1 Terra computer use, and on how an agent logs in without seeing your password, the 1Password for Agents breakdown.

    Frequently asked questions

    What is Gemini Computer Use?

    A Gemini API tool: the model looks at screenshots and returns actions (click, type) that your code runs.

    How much does it cost?

    By my own calculation, about 10 cents per 12-step task with Gemini 3.8 Flash. On January 1 that rate doubles.

    Does Gemini control my computer directly?

    No: it only answers with written actions; your code does the clicking.

    Is it safe to turn loose?

    Not without rules: the documentation warns of possible errors and vulnerabilities, and prompt injection detection is off by default.

    Conclusion

    Gemini Computer Use has arrived as an official tool; what's missing is judgment: where yes, where no, and who checks. Start with a read-only task, with a step cap, and measure before scaling.