Google has introduced Gemini 2.5 Computer Use, a refined version of its Gemini 2.5 Pro model designed specifically for agents that interact directly with user interfaces on the web and mobile platforms. The model demonstrates strong results in web control benchmarks, offering improved speed and lower latency compared to other systems.

Key Capabilities

1. Interface control: The model can perform a range of actions including clicking, typing, scrolling, dragging, and filling forms.
2. Loop execution: Operates within a reliable screenshot–action feedback loop.
3. User confirmation: Prompts users before executing sensitive or high-risk actions.
4. Optimized for browsers: Tuned for web environments with early indications of solid performance on mobile applications.
5. Safety measures: Includes built-in safeguards against malicious use, fraud, and prompt injection attempts.

How It Functions

Gemini 2.5 Computer Use follows a continuous loop process — it captures a screenshot, references prior interactions, predicts the next action, executes it, and repeats the cycle. Its action library includes standard UI behaviors such as typing, scrolling, selecting from dropdown menus, and handling logins. For sensitive operations, it requests explicit user confirmation before proceeding.

Benchmark Results

Testing shows Gemini 2.5 Computer Use outperforming competing models in several key areas:

1. Web task accuracy: Around 72% success rate on live web benchmarks.
2. Complex navigation: Highest scores in multi-step interface tasks.
3. Mobile control: Surpasses earlier baselines on app-based interactions.
4. Latency: Completes tasks in roughly 225 seconds — the fastest among models in its accuracy range.
5. Tester feedback: Reported 50% faster workflows, 18% higher accuracy, and 25% fewer test failures.

Early Applications

Google has already deployed the model internally across projects such as UI testing, Firebase Agent, Project Mariner, and AI Mode. External partners like Poke.com, Autotab, and Google Payments are also using it to enhance automation reliability and system recovery rates.

Availability

Gemini 2.5 Computer Use is now available for public preview through the Gemini API on AI Studio and Vertex AI. Developers can experiment with it in Browserbase or build custom agents using Playwright locally or via cloud-based setups.