Google launches Gemini 2.5 to control interfaces

Google has introduced Gemini 2.5 Computer Use, a model that can operate web pages and applications through their interfaces, as a person would: clicking, typing, scrolling and filling out forms.

The model is available today in public preview for developers through the Gemini API, both in Google AI Studio and Vertex AI. Its goal is to enable agents that can complete digital tasks, even when there is no specific API to connect to a service.

What it can do

Many programs let you automate actions through APIs, which are connections designed for two systems to exchange data. But a large number of tasks still require using buttons, menus and text fields.

Gemini 2.5 Computer Use is designed to handle actions such as:

• Filling out and submitting forms.
• Choosing options in dropdown menus and filters.
• Navigating pages that require signing in.
• Organizing items by dragging them to different areas.
• Scheduling appointments or updating records in web applications.

In one of Google's demonstrations, the agent identifies pets living in California in a form, adds them as customers in a management system for a grooming salon and schedules an appointment with a specialist. In another, it sorts digital notes into the correct categories on a board.

The practical value is clear: you could ask it to move data from one system to another or complete a repetitive task without opening several pages and doing everything manually. The agent does not just generate text. It operates the interface.

How it works

The model receives three elements: the user's request, a screenshot of the environment and the recent history of actions. It then decides what to do and returns an action, usually an instruction to click, type or scroll.

The program integrating it executes that action and sends a new screenshot along with the current URL. The process repeats until the task is complete, an error is found or it stops because of a user decision or a security response.

For sensitive actions, such as making a purchase, the system can ask for confirmation before continuing. Developers can also limit the available actions or add their own functions.

Performance and limits

Google says the model outperforms leading alternatives in several web and mobile control tests, with lower latency, meaning less time between one instruction and the next. The results come from Google's own evaluations, figures reported by Google and tests conducted by Browserbase, so they do not represent a guarantee for every website or application.

The model is optimized mainly for browsers. It also shows promising results on mobile interfaces, but it is not yet optimized to control a computer at the operating system level, such as opening any installed program or freely manipulating the desktop.

The risk of giving it control

An agent that can operate a website can also make mistakes, follow malicious instructions hidden on a page or fall for a scam. Google points to three main risks: intentional misuse to cause harm, unexpected model behavior and prompt injection attacks, in which a website attempts to change the agent's original instructions.

To reduce these risks, Gemini includes security functions in the model and offers additional controls:

• A security service reviews each action before it is executed.
• Developers can require confirmation for high-risk operations.
• The system can block actions such as bypassing CAPTCHAs, compromising a computer's security or controlling medical devices.

Google recommends thoroughly testing any agent before making it available to the public. The preview does not turn these automations into autonomous systems that require no supervision.

Google already uses versions of this technology in interface testing, Project Mariner, Firebase Testing Agent and some AI features in Search. For everyone else, access now begins through the API, with tools such as Playwright or Browserbase's cloud environments.

The important change is the type of task these models can take on: not just answering questions or generating content, but interacting with services you use every day. The next challenge will not only be getting them to click correctly, but ensuring they know when they should not click and ask for your permission in time.