Google has refreshed the Android Bench 2.0 AI development benchmark leaderboard, incorporating test results for several newly added models. Moonshot AI's Kimi K3.1 Pro posted a 24.0% pass rate on long-horizon tasks with an average task completion rate of 82.3%, landing in sixth place. That pass rate is double that of its predecessor, Kimi K3, while test duration and cost have both fallen sharply. DeepSeek V4.1 Flash 1017 recorded a 17.3% pass rate and an average task completion rate of 71.2%, taking ninth place. Its test run took just 12.0 hours at a cost of only $9.9—a standout cost advantage. Each of the two models runs with a different coding agent. The current top three on the leaderboard are Claude Sonnet 5.5, Claude Opus 5.5, and GPT 6.1 Sol. The benchmark targets Android app development scenarios, comprising 30 long-horizon development tasks that span a range of related areas.