Gemini 3.5 Flash Computer Use: Browser Agent Enters Low-Latency Model Layer
Google DeepMind released the computer use capability of Gemini 3.5 Flash on June 24, 2026.
Browser operations are evolving from external script orchestration to native model capabilities, but reliability, security, and cost still determine whether they can be used in production.
Evaluation must be layered by page understanding, action localization, state verification, anomaly recovery, and prompt injection protection; a successful demo alone is insufficient.
Computer use will compress the difference between general RPA and simple browser agents, shifting value to governance, vertical workflows, and reliable execution.
Teams should first select rollbackable, low-permission processes for shadow runs, and use completion rate and human takeover rate to decide scope expansion.
Observe real website success rates, long-task costs, prompt injection incidents, permission controls, and enterprise audit capabilities.