Skills
Google Android Bench: Platform-Specific LLM Leaderboard for Model Selection in Mobile Development
Google released Android Bench, the first official leaderboard evaluating LLMs specifically on Android development tasks sourced from real-world public GitHub repos, covering breaking changes across Android releases, Wear OS networking, and Jetpack Compose migration. Current top scorers are Gemini 3.1 Pro (72.4%), Claude Opus 4.6 (66.6%), and GPT-5.2-Codex (62.5%), with all models completing only 16–72% of tasks — revealing a significant capability gap that general coding benchmarks like SWE-bench don't surface. The dataset and test harness are open-source on GitHub with anti-contamination canary strings, making it the primary benchmark to run before selecting a model for any Android-focused coding agent.
↳ Follow the thread