Google expands Android Bench 2.0 for AI coding agents
Google released Android Bench 2.0 with long-horizon tasks, agent-based evaluation and continuous scoring for complex Android development work. Its initial long-horizon tasks cover dependency upgrades, new features, app creation and cross-platform conversions that may take engineers several days or a week. IT之家 AI and 9to5Google also reported leaderboard updates with reasoning levels, new model results and average invocation costs.
The expanded benchmark evaluates AI coding agents on multistep Android work and combines completion results with invocation cost, giving developers a broader basis for comparing models.

谷歌发布了 Android Bench 2.0—"—谷歌用于评估 Android 开发任务中 AI 模型和智能代理的基准测试框架。此次重大更新引入了长周期任务(LHT)、基于智能体的评估以及连续评分机制,可以更好地评估在复杂的多步骤开发任务中的性能。
Android Bench 于数月前推出",用于通过一系列常见的开发任务对 AI 模型进行评估,其中融入了权限、导航和网络联通性等方面的 Android 最佳实践。随着最新版本的发布,谷歌已经将该基准测试的覆盖范围扩展至更广泛的开发任务。
今天,我们发布了第一批长周期任务(LHT)。这些任务非常复杂,工程师需要花费数天甚至一周的时间才能完成。此外,我们还推出了基于智能体的评估机制,首先采用的是来自相应模型提供商的智能体。
#Google#Android#Android Bench 2.0#Claude Sonnet 5.5#Claude Opus 5.5#GPT-6.1 Sol#Gemini 4 Argon