Android Bench 2.0 - Long-horizon: leaderboard
Google Android Bench 2.0 evaluates agents on 30 long-horizon Android development tasks. Mean task pass rate across repeated runs is higher-is-better and remains separate from the 100-task Android Bench 1.0 cohort. Each row retains its model and agent harness; completion rate is a different diagnostic.
Metric: Mean task pass rate (%) over 30 tasks. Source: developer.android.com. Saturation forecast: Around August 2028. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 5.5 (Max) | 40 | #13 (Claude Sonnet 5.5) |
| 2 | Claude Opus 5.5 (Max) | 37.3 | #3 (Claude Opus 5.5) |
| 3 | GPT-6.1 Sol (Max) | 29.3 | #6 (GPT-6.1 Sol) |
| 4 | GPT-6 Astra (High) | 28 | #5 (GPT-6 Astra) |
| 5 | Claude Fable 5.1 (High) | 22.7 | #9 (Claude Fable 5.1) |
| 6 | Gemini 4 Argon (High) | 21.3 | #8 (Gemini 4 Argon) |
| 7 | Kimi K3 (Max) | 12 | #34 (Kimi K3) |
| 8 | Gemini 3.8 Flash | 8 | #24 |
| 9 | Gemini 3.7 Flash | 7.3 | #22 |
No result here: #10 GPT-6.1 Terra.
Interactive version: theaggregate.ai/benchmark?slug=android-bench-2-0-long-horizon · How It Works · Data refreshed daily, snapshot 2026-10-27.