Android Bench: leaderboard
Version 1.0 of Google's Android Bench: 100 small tasks taken from pull requests in public GitHub Android repositories, such as handling breaking changes between Android releases or migrating to newer Jetpack Compose. An agent built on mini-swe-agent writes a patch, which passes or fails on unit or instrumentation tests. The number is the average percentage of tasks resolved across all runs, higher is better.
Metric: Score (%). Source: developer.android.com. Saturation forecast: Not forecast. 42 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5 | 91.8 | #12 |
| 2 | Claude Fable 5 | 90.9 | #16 |
| 3 | GPT-5.6 Sol | 90.8 | #19 |
| 4 | Kimi K3 | 90.2 | #30 |
| 5 | GPT-5.6 Luna | 87.6 | #68 |
| 6 | Qwen 3.8 Max | 87 | #35 |
| 7 | GPT-5.6 Terra | 86.8 | #41 |
| 8 | GPT-5.5 | 80.2 | #31 |
| 9 | Claude Sonnet 5 | 76.2 | #57 |
| 10 | Gemini 3.6 Flash | 75.6 | #51 |
| 11 | Gemini 3.1 Pro (Preview) | 74.3 | #58 |
| 12 | GPT-5.4 | 74.1 | #79 |
| 13 | Claude Opus 4.8 | 72.4 | #37 |
| 14 | GLM-5.2 | 72.2 | #73 |
| 15 | Gemini 3.5 Flash | 71.1 | #60 |
No result here: #7 GPT-6.1 Sol, #8 Gemini 4 Argon, #10 Muse Wonder Full.
Interactive version: theaggregate.ai/benchmark?slug=android-bench · How It Works · Data refreshed daily, snapshot 2026-10-20.