Overview
- Google announced on June 24, 2026 that the previously standalone computer use capability is now built into the main Gemini 3.5 Flash model to improve performance for agentic tasks.
- Developers and enterprises can access the feature today through the Gemini API and the Gemini Enterprise Agent Platform with reference code and a live demo hosted by Browserbase.
- Example use cases shown by Google include long‑horizon automation such as continuous software testing, feature analysis and automated accessibility audits that rely on screenshots and UI navigation.
- To reduce prompt‑injection risks, Google applied targeted adversarial training and offers two optional enterprise safeguards that require explicit human confirmation for sensitive actions and automatically stop tasks on detected indirect injections.
- Chrome 149 is rolling out a new Select from screen tool that makes it easy to add on‑screen text or images into Gemini prompts, tying the model capability directly to everyday browsing workflows and early user trials.