Gemini 2.5 Brings AI That Clicks, Types, and Navigates Websites

Google is switching on a new kind of AI assistant — one that doesn’t just talk about doing things, it actually does them. The new Gemini 2.5 Computer Use model is now in public preview for developers through the Gemini API in Google AI Studio and Vertex AI. This version gives AI agents the ability to navigate real websites like a human: opening pages, filling out forms, clicking buttons, dragging items, and continuing until the task is done.

The Next Evolution of Today’s AI

Gemini 2.5 Computer Use represents a significant leap forward in AI interaction. Instead of relying solely on structured APIs, this model uses a more visual and mechanical approach. The process runs in a loop: your code sends a screenshot of the current browser screen and recent actions to the model. Gemini analyzes the visual data and replies with a function call such as “click,” “type,” or “scroll.” The client executes the command, then sends back a new screenshot and URL. The loop continues until the task completes or a safety stop triggers.

This method effectively gives the AI “hands and eyes” for the web. It can work behind login screens, interact with dropdowns, and even navigate dashboards that lack public APIs. In short, Gemini can now handle interfaces built for humans — not just machines.

Browser Control Comes First

Google says this model is optimized primarily for web browsers, with promising early performance on mobile interfaces. Full desktop-level control isn’t the focus yet, but the progress is clear. In recent benchmark tests like Online-Mind2Web and WebVoyager, Gemini 2.5 Computer Use led the pack, offering both higher success rates and lower latency in Browserbase’s environment.

This makes it a potential game-changer for workflows that depend on dynamic web navigation — for instance, managing account dashboards, scheduling bookings, or automating research. It’s the first real step toward AI that performs complex browser tasks instead of merely describing them.

Safety Comes First

Google emphasizes that safety is a built-in layer, not an afterthought. Every AI action can be routed through a step-by-step safety service before it executes. Developers can also configure user confirmations for sensitive tasks such as purchases or operations that might affect system integrity. Additionally, teams can restrict the set of allowed actions to prevent unwanted clicks or inputs.

Still, Google cautions developers to perform extensive testing before deploying production agents. As these AI models gain more autonomy, ensuring predictable behavior is critical to maintaining trust and security.

Developer Access and Integration

To help developers experiment, Google has released several tools and demos. You can try the hosted experience via Browserbase, review sample agent loops, or explore documentation for local testing with Playwright. The setup allows for both quick experimentation and deeper integration into automated systems or browser-based workflows.

If parts of this technology feel familiar, they should. Variants of Gemini’s Computer Use capabilities have already been active behind the scenes in Project Mariner, Firebase Testing Agent, and even experimental features within Search’s AI Mode. The difference now is that the preview opens the door to public development — a clear sign that Google sees this as a core part of Gemini’s evolution.

From Assistant That Suggests to Assistant That Acts

With the launch of this public preview, Gemini is shifting from being a voice of advice to an AI that takes real action. It no longer just recommends what to click — it can actually click it for you. For users and developers who live and work on the web, that’s a powerful leap forward. It means the start of automated workflows that combine reasoning, perception, and physical-like web interaction.

In 2025, Gemini isn’t just another chatbot. It’s becoming a true digital operator — one capable of executing browser tasks end-to-end. For anyone whose workflow revolves around web interfaces, Gemini 2.5 might be the most impactful tool Google has released this year.