In detail
- The model can now control browsers, mobile devices, and desktop environments independently—previously available only via a separate Gemini 2.5 model.
- On the OSWorld benchmark, Gemini 3.5 Flash scores 78.4, beating GPT-5.4 mini (72.1) but trailing Anthropic Opus 4.8 (83.4).
- Google uses adversarial training and optional enterprise safeguards against prompt injection; sandboxing and human oversight are recommended.
Why it matters
For businesses planning automation of office workflows, software testing, or data processing, direct screen control by AI agents becomes a key differentiator—saving development time and unlocking new use cases.
For you Explore how you could use Gemini 3.5 Flash with Computer Use to automate repetitive tasks in your systems—especially for RPA-like scenarios.