Gemini 3.5 Flash gets computer use, at near-frontier accuracy

Google has folded computer use into Gemini 3.5 Flash, its fast and cheaper model tier, as a built-in tool rather than a separate model. The capability lets an agent see a screen, reason about it, and act across browser, mobile, and desktop, which is the basic loop behind tasks like filling forms, testing software, or working through a web application the way a person would. Until now this lived in a standalone Gemini 2.5 computer use model.

The number that makes this interesting is the benchmark. On OSWorld-Verified, which measures how well an agent completes real computer tasks, Gemini 3.5 Flash scores 78.4 percent. That is a large jump from the 65.1 percent of the previous Flash model, and it sits within a point of the frontier: GPT-5.5 leads at 78.7 percent and Claude Opus 4.7 is at 78.0 percent. A cheap, fast model is now doing computer-use work at close to the level of the most expensive ones.

Google is also explicit about the risk that comes with letting a model click around on your behalf. The model is trained against adversarial inputs, and enterprises can turn on safeguards: asking a user to confirm sensitive actions, and stopping a task automatically when the system detects an indirect prompt injection. Developers can try it through the Gemini API and a Browserbase demo, described in Google's announcement.

Why it matters

If you have wanted to build computer-use agents but balked at the cost of running a frontier model in a loop, a Flash-tier model at 78.4 percent changes the math. Test it on your own tasks, and turn on the prompt-injection safeguards before you point it at anything that can spend money or touch real data.