XNews

Google Launches Android Bench 2.0 to Measure Model-Agent Performance on Real Android Tasks

Google Launches Android Bench 2.0 to Measure Model-Agent Performance on Real Android Tasks

Google introduced Android Bench 2.0, a benchmark for evaluating model-agent combinations on real Android development tasks from production apps. The benchmark includes 30 tasks across app conversion, architecture migration, and new feature development. Top performers were GPT-6 Astra+Codex and Claude Fable 5.1+Claude Code, with pass rates up to 28%.

Source posts