Overview
- Android Bench moved to the Harbor framework and an updated benchmarking agent in the July release to make tests easier to run, share, and reproduce by third parties.
- Google re-ran every model under the new Harbor-based agent to set an updated baseline, producing minor score shifts while keeping historical results archived and accessible.
- The leaderboard was expanded with eight new entries and refreshed rankings that place Claude Fable 5 first overall and GLM 5.2 first among open-weight models.
- The benchmark now reports capability as well as cost and efficiency so developers can weigh performance against latency and pricing when choosing models for Android tasks.
- Android Bench is open for community contributions on GitHub and Harbor Hub so developers can submit Android-specific tasks and run or share their own evaluations to shape the dataset.