# Gemini 3.5 Flash gained built-in computer use > Google integrated screen-based browser, mobile and desktop interaction into its main Flash model for custom automation agents. Canonical URL: https://www.devobs.io/news/news-gemini-3-5-flash-built-in-computer-use/ By: Finn Larson Published: 2026-09-06T11:58:54.628Z Updated: 2026-09-06T11:58:54.628Z Event date: 2026-06-24 Section: Developer tools Google added computer use as a built-in tool in Gemini 3.5 Flash. The [June 24 announcement](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-computer-use-gemini-3-5-flash/) says the capability can see and act across browser, mobile and desktop environments. It had previously been offered through a standalone Gemini 2.5 computer-use model; the new release places it alongside Flash's function calling and built-in grounding tools. ## Screen actions create a different integration contract An API call usually has a typed schema and a predictable error response. A screen action depends on layout, focus, timing, localization and visual state. The same click can mean something else after a page update or delayed animation. Developers need to model observation and action as separate steps and confirm that the target state still holds immediately before a consequential input. Computer use can cover software without an API, but it should not inherit broad desktop credentials by default. Run agents in isolated profiles or virtual environments and expose only the applications and accounts required for the task. ## Test recovery, not just completion Evaluation should include pop-ups, expired sessions, slow rendering, duplicate submissions, reordered controls and a page that returns untrusted instructions. Record screenshots or structured accessibility state around each action while redacting secrets. Define a step limit and stop when the observed state diverges from the expected one. For write operations, require an application-side policy or human checkpoint based on the concrete target and effect. Make repeated actions idempotent when possible. Google's integration reduces the need to combine a separate model with Flash, but the reliability boundary remains the changing interface the agent is driving. ## Source references -