ErdosBench: leaderboard
Ulam's ErdosBench evaluates research-level, Erdős-inspired mathematical problem solving on 226 problems — rewarding decisive obstructions, correct theorem use, and scoped partial progress while penalizing unsafe solved claims; scored by external-judge grade average (0-4) with model-specific proof audits.
Metric: Judge Average (0-4). Source: www.ulam.ai. Saturation forecast: Around January 2028. 19 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Fable 5.1 (xHigh) | 3.25 | #9 (Claude Fable 5.1) |
| 2 | Claude Opus 5.5 (xHigh) | 3.24 | #3 (Claude Opus 5.5) |
| 3 | GPT-6 Astra (xHigh) | 3.23 | #6 (GPT-6 Astra) |
| 4 | GPT-5.6 Sol (xHigh) | 3.12 | #19 (GPT-5.6 Sol) |
| 5 | Muse Wonder | 3.05 | #15 |
| 6 | Gemini 4 Argon | 3.04 | #8 |
| 7 | MiMo-V2.6-Pro | 3.02 | #42 |
| 8 | GLM-5.3 Flash | 2.95 | #65 |
| 9 | Muse Spark 1.3 | 2.86 | #24 |
| 10 | Kimi K3 (Max) | 2.74 | #30 (Kimi K3) |
| 11 | Muse Spark 1.2 | 2.73 | #62 |
| 12 | GPT-5.5 (xHigh) | 2.68 | #31 (GPT-5.5) |
| 13 | Kimi K2.7 Code | 2.62 | #89 |
| 14 | Claude Opus 4.8 (Max) | 2.51 | #37 (Claude Opus 4.8) |
| 15 | GLM-5.2 | 2.4 | #73 |
No result here: #7 GPT-6.1 Sol, #10 Muse Wonder Full.
Interactive version: theaggregate.ai/benchmark?slug=erdosbench · How It Works · Data refreshed daily, snapshot 2026-10-20.