Google releases Android Bench 2.0 for agent evaluation

Android Bench 2.0 tests model-agent pairs on real Android tasks; highest pass rate 28%. GPT-6 Astra+Codex and Claude Fable 5.1+Claude Code lead; most cross-platform migrations failed.

Android Bench 2.0 tests model-agent pairs on real Android tasks; highest pass rate 28%. GPT-6 Astra+Codex and Claude Fable 5.1+Claude Code lead; most cross-platform migrations failed.