SelfBench - calcom/cal.diy: leaderboard
Coding agents implement changes reconstructed from 39 merged pull requests of calcom/cal.diy. Hidden tests decide task success; seven model, harness and reasoning configurations are evaluated separately, including the Dari routing system. The publisher supplies tasks and trials. The score is passed tasks divided by 39 as a percentage, higher is better. Source: selfbench.dev, release f4dd2825-e652-4f52-8535-098e0ce3ad96.
Metric: Task success (%) on 39 tasks rebuilt from the repository's merged pull requests (hidden tests). Source: selfbench.dev. Saturation forecast: Around February 2028. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5.5 (Max) | 74.36 | #3 (Claude Opus 5.5) |
| 2 | GPT-6.1 Sol (Max) | 69.23 | #6 (GPT-6.1 Sol) |
| 3 | GPT-6.1 Sol (Medium) | 69.23 | #6 (GPT-6.1 Sol) |
| 4 | GPT-6 Luna (Max) | 53.85 | #60 (GPT-6 Luna) |
| 5 | GPT-6 Luna (Medium) | 48.72 | #60 (GPT-6 Luna) |
No result here: #5 GPT-6 Astra, #8 Gemini 4 Argon, #9 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=selfbench-calcom-cal-diy · How It Works · Data refreshed daily, snapshot 2026-10-27.