AndroidLife: Can an AI agent survive a day in the life of a real user? Qwen3.8-27b run: 56.7% SR

The AndroidLife benchmark tested the Qwen3.8-27b AI model's ability to perform 60 real-world tasks on a OnePlus phone, achieving a 56.7% success rate. The model struggled with multi-application tasks and instances requiring user interaction, failing 43% of the tasks. During the test, the phone's chip reached a peak temperature of 98.2 C, and 69% of the battery was consumed. This benchmark aims to evaluate AI agents in a realistic mobile environment, tracking performance, thermals, and battery usage.