DoGBench: leaderboard
Agents write or update user-facing software documentation in response to real repository events in open-source projects such as Helm and PostHog, including cases where the right answer is to change nothing. Patches are graded against rubrics validated with project maintainers and combined into a 0-100 composite. Each row is one model in one coding harness. Held-out split, higher is better. Source: dogbench.ai.
Metric: Composite score (0-100), held-out split (117 items). Source: dogbench.ai. Saturation forecast: Around July 2028. 7 models tracked. Updated 3 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.8 Max | 47.26 | #64 |
| 2 | GPT-5.6 Sol | 46.22 | #33 |
| 3 | GLM-5.2 | 44.16 | #152 |
| 4 | Kimi K2.7 Code | 43.88 | #199 |
| 5 | Claude Opus 4.8 | 41.16 | #62 |
| 6 | Claude Sonnet 4.6 | 40.02 | #173 |
| 7 | GPT-5.5 | 34.02 | #50 |
No result here: #1 Claude Fable 5.2, #2 GPT-6.2 Astra, #3 Gemini 4 Carbon.
Interactive version: theaggregate.ai/benchmark?slug=dogbench · How It Works · Data refreshed daily, snapshot 2026-11-17.