Updated
Updated · Ars Technica · Jul 8
Google Adds 8 Models to Android Bench as Gemini 3.1 Pro Slips to No. 5
Updated
Updated · Ars Technica · Jul 8

Google Adds 8 Models to Android Bench as Gemini 3.1 Pro Slips to No. 5

2 articles · Updated · Ars Technica · Jul 8

Summary

  • Google expanded Android Bench with eight new LLMs, a revised framework and new cost and efficiency metrics for its 100-task Android app development benchmark.
  • Claude Fable 5 took a clear lead at 84.5% accuracy, while Gemini 3.1 Pro ranked fifth behind GPT 5.4, Claude Sonnet 5 and Claude Fable 5.
  • The update broadens the leaderboard to include open-weight models and newer flagship systems such as Claude Opus 4.8, GLM 5.2, Kimi K2.7 Code, MiniMax M3 and Qwen 3.7 variants.
  • Google said developers can now run their own tests and submit feedback, positioning Android Bench as a more practical tool for comparing coding agents as code generation gains traction.

Insights

Google's AI ranks fifth on its own benchmark. Is the top-ranked model actually worth the cost for developers?
As AI models master Android code, can we trust benchmarks that might just be testing their memory?
Beyond accuracy, what hidden costs and efficiency trade-offs are AI coding assistants hiding from Android developers?