Android Bench 2.0 Shows Top Models Fail 28% Pass Rate on Real-World Android Tasks

Google released Android Bench 2.0, a benchmark evaluating model+agent pairs (e.g., GPT+Codex, Claude+Claude Code) on real production Android apps. Top models achieved a maximum 28% pass rate, with all models scoring 0/5 on cross-platform migration tasks (Flutter/React Native to native Android) and most failing XML-to-Compose migrations.