Gemini 3.5 Flash gets built-in computer use
Google folded computer use into Gemini 3.5 Flash on June 24, scoring 78.4 on OSWorld and shipping two guardrails against agents that click on your behalf.
Google made computer use a built-in tool in Gemini 3.5 Flash on June 24, letting developers build agents that see and operate a screen across browsers, mobile, and desktop. The model scored 78.4 on OSWorld, the standard test for computer-use agents, up from 65.1 for the previous Gemini 3 Flash and just behind GPT-5.5 at 78.7. Anthropic’s Opus 4.8 still leads at 83.4. Like every entry on that board, the score is self-reported.
The capability used to be a standalone feature in Gemini 2.5. Folding it into the cheaper Flash model aims it at long-running jobs like continuous software testing and routine work across web apps. Developers can reach it now in the Gemini API, and it also ships in Google’s Enterprise Agent Platform, with a hosted demo on Browserbase and a reference implementation on GitHub.
Google also shipped two safeguards, because an agent that clicks and types for you is a new attack surface. One requires explicit user confirmation before sensitive actions. The other halts a task automatically if it detects indirect prompt injection, the trick where hidden instructions in a page or email hijack the agent. Google says it did targeted adversarial training too, and still recommends sandboxing, human review, and tight access controls on top.
What it means for you
Cheap screen automation is now realistic for small teams, but the guardrails are the headline. Before you let one of these agents loose, read up on how an email can hijack your AI and how to keep a coding agent from running attacker code. The same caution that applies to agents handed their own browser sessions applies here.