Google bakes computer control directly into Gemini 3.5 Flash, letting the model see and operate your screen

Insight summary

• •Google integrated 'Computer Use' into Gemini 3.5 Flash to enable the model to see, understand, and interact with computers and devices directly.
• •This capability allows agents to operate across browser, mobile, and desktop environments for tasks like software testing and office automation.
• •Gemini 3.5 Flash scored 78.4 on the OSWorld benchmark, outperforming earlier Gemini versions and GPT-5.4 mini but slightly behind GPT-5.5 and Anthropic's Opus 4.8.
• •To prevent prompt injection attacks, Google employs adversarial training and two enterprise safeguards requiring user confirmation or automatic task stoppage.
• •Additional security recommendations include sandboxing, human oversight, and strict access controls, detailed in Google's best practices documentation.
• •The new feature is accessible via the Gemini API and the Gemini Enterprise Agent Platform, with a Browserbase demo and GitHub reference implementation provided.
• •This integration replaces the previously separate Gemini 2.5 model's computer interaction capabilities.