Google launches Android Bench 2.0 for AI coding agents; top pass rate 28%

Google introduced Android Bench 2.0, a benchmark that evaluates model-plus-agent pairs on real Android engineering tasks based on Google best-practice documents and production apps. In the new 30-task track, GPT 6 Astra with Codex and Claude Fable 5.1 with Claude Code led, but the maximum pass rate was 28%. No model passed cross-platform conversions or most XML-to-Compose migrations.