Google Gemini 3.5 Flash gains built-in computer use control

Letting an AI “drive” your computer sounds like a novelty, but it’s a big step toward agentic systems that operate end‑to‑end in real software environments. By baking computer‑use directly into Gemini 3.5 Flash, Google is turning what used to be a specialized model into a default capability: the agent can now see the UI, click, type, navigate and mix that with traditional tool calls, all from one model endpoint.

This matters for the race to AGI because control, not just prediction, is where a lot of value – and risk – concentrates. Agents that can reliably execute long‑horizon workflows across browsers and enterprise apps are much closer to digital workers than chatbots. Google’s move also directly targets OpenAI’s push into agents and undercuts one of GPT‑5.5’s differentiation points by offering comparable OSWorld‑style performance at lower per‑token prices.

On the flip side, first‑class computer use intensifies safety questions. Prompt‑injection, data exfiltration and abuse of credentials become far more consequential when the model can move through your actual systems rather than a sandbox. The labs that can prove they’ve tamed those risks while still shipping usable agents will be the ones that define how quickly agentic AI escapes the lab and becomes day‑to‑day infrastructure.