Overview

• Developers can access the model now via the Gemini API in Google AI Studio and Vertex AI, with a public demo available on Browserbase.
• Gemini 2.5 Computer Use powers agents that click, type, scroll, drag and drop, manipulate forms and dropdowns, and work behind logins using a cyclical tool loop of screenshots and action history.
• Google positions the model as optimized for web browsers with promising mobile results, noting it is not yet optimized for desktop OS-level control.
• Google reports superior accuracy and lower latency versus leading alternatives on multiple web and mobile control benchmarks, citing evaluations including Browserbase tests.
• Safety measures include model-level training, a per-step inference-time safety service, and developer-set instructions to refuse or require confirmation for high-risk actions such as purchases or CAPTCHA bypass attempts.