Overview

• The specialized variant of Gemini 2.5 Pro lets agents operate websites by clicking, typing, scrolling, dragging and dropping, and submitting forms.
• A new computer_use tool runs in a loop from screenshot and action history to a function‑call action, with the client executing steps and returning an updated view.
• Google says it outperforms leading alternatives on web and mobile control benchmarks with lower latency, based on self‑reported and Browserbase evaluations.
• Safety features include an out‑of‑model per‑step review and developer rules that can require confirmation or block high‑risk actions such as purchases or CAPTCHA bypasses.
• Access is available now through the Gemini API in Google AI Studio and Vertex AI, with a live demo on Browserbase and internal use reported in Mariner, Firebase Testing Agent, and AI Mode in Search.