Google launched a new tool Tuesday that lets its Gemini AI model take control of web browsers to handle everyday online tasks. The Gemini 2.5 Computer Use feature aims to make AI agents more useful for developers and businesses.
This release marks a key step in AI development. It allows companies to automate routine work that relies on user interfaces, such as filling out forms or gathering data from sites without APIs. Businesses in e-commerce, testing, and customer service stand to gain the most. The tool could cut down on manual labor and speed up processes, though experts note it is still in early stages.
Model Mimics Human Web Actions
The Gemini 2.5 Computer Use model builds on the Gemini 2.5 Pro foundation, released June 17, 2025. It uses visual reasoning to scan screenshots of a browser screen and decide on next steps, like where to click or what to type. Developers feed it a user request, a current screenshot, and past actions. The AI then outputs commands for up to 13 UI moves, including scrolling down a page or dragging items.
In one demo, the model signed up fictional pets for a spa visit by pulling details from a site and booking an appointment for October 10. Another example sorted sticky notes on a web app into categories for an art club event. These tasks show how the tool handles real-world workflows behind login walls.
Google says the model shines on web browsers but holds promise for mobile apps too. It does not yet control full desktop systems. Benchmarks show it leads competitors. On the Online-Mind2Web test, it scored 65.7 percent, topping Claude Sonnet 4 at 61 percent and OpenAI’s agent at 44.3 percent.
Developers Build Agents with New API Tools
The feature rolled out in public preview via the Gemini API on Google AI Studio and Vertex AI. Developers must code a loop to run actions in a sandboxed browser, capture new screenshots, and loop back to the model until the job ends. Tools like Playwright help execute clicks at exact spots, scaled from 0 to 999 coordinates to fit any screen.
No free tier exists for this paid model, which charges $1.25 to $2.50 per million input tokens and $10 to $15 per million output tokens, matching Gemini 2.5 Pro rates. Early users include Google’s own teams for UI testing and payments checks, where it fixed over 60 percent of failed runs.
Sundar Pichai, Google CEO, called it a big leap. “These are early days, but the model’s ability to interact with the web – like scrolling, filling forms and navigating dropdowns – is an important next step in building general-purpose agents.”
Safety Checks Limit AI Risks
Google baked safeguards into the model to curb misuse. A per-step review flags risky moves, like solving CAPTCHAs or buying items, and demands user okay before going ahead. Developers can add rules to block actions on certain sites or pause for human input on deals or data shares.
The setup runs in secure containers to avoid broader harm. Still, the preview stage means errors may happen, such as wrong clicks from bad screenshots. Experts urge firms to log all steps for audits.
This tool fits into wider AI trends. Rivals like OpenAI and Anthropic offer similar controls, but Gemini edges them in speed and web scores. Some worry about privacy in automated browsing, though no rules target this yet.
Path Ahead for AI Automation
The preview stays open now with no set end date. Google plans wider tests for personal aides and more app types. Firms can try demos on Browserbase to see it play games or scan news.
For business readers, this means easier ways to build smart tools. Watch for updates on full launch and new benchmarks in coming months.
