Google publishes Android Bench results, with Kimi K3.1 Pro at 24.0 percent and DeepSeek V4.1 Flash at 9.9 dollars per run
Google Android Developers published the first Android Bench results. Kimi K3.1 Pro achieved a 24.0 percent pass rate over 30 long-horizon tasks with 27.4 hours average time and 317.4 dollars per run, while DeepSeek V4.1 Flash reached 17.3 percent with 12 hours and 9.9 dollars per run.
Google's official Android Developers account posted on the social platform on November 10, publishing for the first time the benchmark scores of OpenAI GPT-6.1 Terra, Kimi K3.1 Pro and DeepSeek V4.1 Flash on Android Bench. Android Bench is an AI coding benchmark Google introduced for Android development scenarios, focusing on whether models can complete real engineering tasks such as application creation, feature development, code migration and application conversion. Kimi K3.1 Pro, which participated using the kimi-code agent, achieved a 24.0 percent pass rate on 30 long-horizon tasks with an average completion rate of 82.3 percent, an average duration of 27.4 hours and a benchmark run cost of 317.4 dollars per run. In horizontal comparison, Kimi K3.1 Pro scored above GPT 6 Astra, Claude Fable 5.1 and Gemini 4 Argon but below GPT 6.1 Terra at 28.7 percent. DeepSeek V4.1 Flash 1017 entered through the opencode agent, reaching a 17.3 percent pass rate, 71.2 percent average completion, 12.0 hours average time and 9.9 dollars per benchmark run. Although its pass rate is below that of Kimi K3.1 Pro, its advantages in speed and cost are clear. This round of results put Chinese models on the same stage as OpenAI and Anthropic frontier models on Google's own Android development benchmark for the first time.