MathDuels - Author Rating: leaderboard

MathDuels (2026) is a self-play mathematics benchmark: each model authors 30 hardened problems and solves those the other models wrote, with symbolic answer checking and a verifier removing ill-posed problems; ratings come from a Rasch model on an Elo-style scale. This board is the author rating, the mean rated difficulty of the valid problems the model wrote, in rating points on an open scale; higher is better.

Metric: Author rating (Elo-style points, open scale): mean rated difficulty of the model's own valid problems, capped for problems whose author-supplied answer is wrong; 19 models each author 30 hardened problems (six mathematical domains) and solve every problem authored by the others; answers checked symbolically, ill-posed problems removed by a GPT-5.4-high verifier; Rasch model on a shared rating scale anchored at Grok-4.1-fast-high = 1500, k = 30; higher is better. Source: arxiv.org. Saturation forecast: Around May 2027. 15 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Gemini 3.1 Pro (Preview)High1624#99
2GPT-5.4High1473#135
3GPT-5.2High1427#185
4Gemini 3 FlashHigh1401#175
5Claude Opus 4.6High1307#113
6Claude Sonnet 4.6High1254#173
7GPT-5.4 MiniHigh1218#406
8Grok 4.1 FastHigh1169#427
9Kimi K2.51158#266
10Qwen 3.5 397B A17B1123#264
11MiniMax-M2.71122#539
12GLM-51095#269
13DeepSeek V3.21072#419
14Step 3.5 Flash1032#526
15Grok 4.20High1020#301

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=mathduels-author-rating · How It Works · Data refreshed daily, snapshot 2026-11-21.