Google adds native computer use to Gemini 3.5 Flash

Google DeepMind announced that Gemini 3.5 Flash now supports native computer use as a built-in tool, enabling custom agents that can see and take action across browser, mobile and desktop interfaces. Rather than wiring each application through a bespoke API, the model works visually: it captures screenshots, interprets graphical user interfaces, and performs actions like clicking buttons, filling forms and operating apps. DeepMind says the feature posts strong agentic and UI-control benchmark results and ships with optional enterprise safeguards to reduce risk.

The technical bet is that GUI-level operation generalizes better than API-by-API integration — an agent that can read a screen can, in principle, drive any software a human can, including legacy apps with no programmatic interface. Putting this in the cheaper, faster Flash tier (rather than only a premium Pro model) signals Google wants high-volume agentic automation to be economically viable.

Competitively, this is Google's answer to a category Anthropic pioneered with Claude's 'computer use' and that OpenAI and others have pushed via operator-style agents. Doing it natively in Flash, and pricing it for scale, is the differentiator Google is leaning on. It also arrives awkwardly alongside news that the larger Gemini 3.5 Pro slipped to July, so Flash's computer-use launch is partly carrying the momentum.

Caveats: vision-driven GUI agents remain brittle on dynamic layouts, slow relative to API calls, and pose real security risks (prompt injection via on-screen content, unintended clicks). The optional 'enterprise safeguards' framing suggests Google knows reliability and safety are the gating factors. Watch for independent benchmarks, latency/cost data, and how the safeguards constrain high-risk actions like payments or system settings.