Gemini 3.5 Flash sees your screen, but does it beat GPT-5.5 on OSWorld?
google integrated computer use into gemini 3.5 flash, achieving a 78.4 osworld score that competes with gpt-5.5. this automates browser and desktop tasks, raising the stakes for autonomous agent safety against prompt injections.
google upgraded gemini 3.5 flash with native computer use capabilities, allowing developers to build agents that autonomously operate across web browsers, mobile apps, and desktop environments using tools like function calls and maps. the model achieves a 78.4 score on the osworld benchmark, narrowly trailing gpt-5.5 at 78.7 and anthropic's opus 4.8 at 83.4, but beating its predecessors and gpt-5.4 mini. to address security risks inherent in autonomous computer control, google implemented adversarial training and enterprise features like user confirmation requirements and automatic task stops upon detecting indirect prompt injections. this move accelerates the race for agentic workflows, shifting focus from model reasoning to practical operational capability and safety.