SelfBench - dubinc/dub.link: leaderboard
Coding agents implement changes reconstructed from 21 merged pull requests of the TypeScript repository dubinc/dub.link. Hidden tests decide task success; nine model, harness and reasoning configurations are evaluated separately, in Claude Code, Codex and pi, including the Dari routing system. The publisher, mupt-ai, supplies tasks and trials. The score is passed tasks divided by 21 as a percentage, higher is better. Source: selfbench.dev, release 6682a567-8f7c-4221-8a3a-be4f6ad72264.
Metric: Task success (%) on 21 tasks rebuilt from the repository's merged pull requests (hidden tests). Source: selfbench.dev. Saturation forecast: Around February 2027. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5.5 (High) | 80.95 | #3 (Claude Opus 5.5) |
| 2 | GPT-6.1 Sol (High) | 76.19 | #7 (GPT-6.1 Sol) |
| 3 | Claude Opus 5.5 (Low) | 71.43 | #3 (Claude Opus 5.5) |
| 4 | Claude Haiku 5.5 (High) | 66.67 | #43 (Claude Haiku 5.5) |
| 5 | Claude Haiku 5.5 (Low) | 57.14 | #43 (Claude Haiku 5.5) |
| 6 | GPT-6 Luna (High) | 42.86 | #61 (GPT-6 Luna) |
| 7 | GPT-6 Luna (Low) | 4.76 | #61 (GPT-6 Luna) |
No result here: #6 GPT-6 Astra, #9 Gemini 4 Argon, #10 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=selfbench-dubinc-dub-link · How It Works · Data refreshed daily, snapshot 2026-11-05.