PanCanBench - Human Curated Rubrics: leaderboard
Figure5 reproduction: same6203eligible saved responses across22models, ClaudeOpus5 judge, final human3639-item/282question rubric and separate synthetic rubric. Mean per-question percentage includes negative penalties and genuinezeros, excludes explicit NO RESPONSE GENERATED. Synthetic35interpretation/direction flags remain unadjudicated; no manual score changes. Not pooled-points percentage or the earlier3130-item manuscript draft. Human Curated Rubrics. Higher is better; negative percentages permitted by rubric penalties.
Metric: Mean rubric score (%). Source: github.com. Saturation forecast: Not forecast. 22 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | O3 | 82.19 | #228 |
| 2 | Grok 4 | 81.59 | #307 |
| 3 | GPT-5 | 78.67 | #167 |
| 4 | Gemini 2.5 Flash | 75.54 | #453 |
| 5 | O4 Mini | 72.39 | #353 |
| 6 | Gemini 2.5 Pro | 71.86 | #266 |
| 7 | Gemma 3 27B (IT) | 70.49 | #1289 |
| 8 | Gemma 3 12B (IT) | 70.22 | #1837 |
| 9 | GPT-4.1 | 69.87 | #495 |
| 10 | OLMo 3 32B (Thinking) | 68.58 | #1884 (OLMo 3 32B) |
| 11 | Qwen 3 32B | 66.24 | #960 |
| 12 | OLMo 3.1 32B Instruct | 62.88 | #2384 |
| 13 | Qwen 3 8B | 60.99 | #1830 |
| 14 | Qwen 3 14B | 59.72 | #1233 |
| 15 | Claude Opus 4.1 (20250805) | 54.97 | #257 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=pancanbench-human-curated-rubrics · How It Works · Data refreshed daily, snapshot 2026-11-13.