ComBench: leaderboard
ComBench evaluates rigorous proof reasoning and constructive realization on 100 Olympiad-level combinatorics problems (50 analysis-centric, 50 construction-centric). This board reports Overall average normalized score (%) over four sampled solutions across the 100 problems (proof rubric score for analysis items, verifier-gated score for construction items); higher is better.
Metric: Overall average normalized score (%) over four sampled solutions across the 100 problems (proof rubric score for analysis items, verifier-gated score for construction items); higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 65.4 | #49 |
| 2 | Gemini 3.1 Pro (Preview) (Thinking) | 60.3 | #94 (Gemini 3.1 Pro (Preview)) |
| 3 | Kimi K2.6 | 53.5 | #178 |
| 4 | DeepSeek V4 Pro | 45.2 | #188 |
| 5 | Qwen 3.6 Max Preview | 24.9 | #159 |
| 6 | GLM-5.1 | 23.6 | #179 |
| 7 | Qwen 3.6 35B A3B | 20.3 | #462 |
| 8 | Nemotron Cascade 2 30B A3B | 19.6 | #1091 |
| 9 | Gemma 4 31B (IT) | 16.8 | #342 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=combench · How It Works · Data refreshed daily, snapshot 2026-11-13.