Google Launches Android Bench 2.0 for Real Android Dev Agent Evaluation

Google released Android Bench 2.0, evaluating model+agent combinations on real Android tasks from Google best practices. Top models achieved a 28% pass rate, with GPT-6 Astra and Claude Fable leading; Google's Gemini models placed lower but offered cost advantages. Many tasks, especially XML-to-Compose migration and cross-platform conversion, had zero passes across all models.