Google Android Bench 2.0 evaluates model-Agent combos on real Android dev tasks

Google launched Android Bench 2.0, measuring model-plus-Agent systems on 30 real Android development tasks including app conversion, architecture migration, and new feature development. Top scorers GPT-6 Astra and Claude Fable reached 28% pass rate; all models failed on cross-platform migration and most XML-to-Compose conversions.