Google launches Android Bench 2.0 for real‑world agentic coding tasks

Google released Android Bench 2.0, a benchmark testing AI model + agent combinations on real Android engineering tasks from production apps. Top pass rate was 28%, with GPT‑6 Astra+Codex and Claude Fable+Claude Code leading, while several tasks (e.g., Flutter→Android migration) had 0/5 passes across all models.