Math
ProofBench Leaderboard
ProofBench evaluates a model's ability to write and check rigorous mathematical proofs, rather than just produce a final numeric answer. It probes step-by-step logical correctness, which is a harder and more distinct skill than solving for a number.
Source: epoch26 open models ranked+56 proprietaryData through Oct 2026
All models ranked on ProofBench
Proprietary / closed models are shown dimmed — you can't run them locally, but they show where the open field stands.
| # | Model | Score | VRAM @ Q4 |
|---|---|---|---|
| 1 | Claude Fable 5.1 Max · proprietary | 100.0% | API only |
| 2 | Claude Opus 5.5 Max · proprietary | 100.0% | API only |
| 3 | Claude Sonnet 5.5 Max · proprietary | 100.0% | API only |
| 4 | Claude Opus 5 Max · proprietary | 99.0% | API only |
| 5 | Gemini 4 Argon (unspecified) · proprietary | 99.0% | API only |
| 6 | GPT 6 Astra (unspecified) · proprietary | 99.0% | API only |
| 7 | GPT 6.1 Sol (unspecified) · proprietary | 99.0% | API only |
| 8 | Claude Fable 5 Max · proprietary | 95.0% | API only |
| 9 | Claude Haiku 5.5 Max · proprietary | 89.0% | API only |
| 10 | Kimi K3 · 2779.9B | 87.0% | ~1674 GB |
| 11 | Muse Wonder 1.0 (unspecified) · proprietary | 84.0% | API only |
| 12 | GPT 5.6 Sol Max · proprietary | 83.0% | API only |
| 13 | GPT 6 Sol (unspecified) · proprietary | 83.0% | API only |
| 14 | Claude Sonnet 5 Max · proprietary | 77.0% | API only |
| 15 | Hy4 Preview · 780.0B | 75.0% | ~1561 GB |
| 16 | GPT 5.6 Terra (xhigh) · proprietary | 74.0% | API only |
| 17 | MiMo V2.6 Pro RL · 1020B | 70.0% | ~613 GB |
| 18 | Claude Opus 4.8 Max · proprietary | 69.0% | API only |
| 19 | GPT 6 Luna (unspecified) · proprietary | 64.0% | API only |
| 20 | MiMo V2.6 Flash RL · 309B | 63.0% | ~186 GB |
| 21 | GPT 5.6 Luna Max · proprietary | 60.0% | API only |
| 22 | Gemini 3.7 Flash (unspecified) · proprietary | 58.0% | API only |
| 23 | Muse Spark 1.3 Max · proprietary | 58.0% | API only |
| 24 | Qwen3.8 Max (unspecified) · proprietary | 58.0% | API only |
| 25 | DeepSeek V4 Flash 0731 · 304.2B | 56.0% | ~183 GB |
| 26 | GPT 5.4 (Mar 05, 2026, xhigh) · proprietary | 56.0% | API only |
| 27 | Muse Spark 1.3 (unspecified) · proprietary | 55.0% | API only |
| 28 | Claude Opus 4.7 Max · proprietary | 54.0% | API only |
| 29 | DeepSeek V4.1 Flash · 763.2B | 54.0% | ~458 GB |
| 30 | Grok 4.6 (unspecified) · proprietary | 51.0% | API only |
| 31 | Claude Opus 4.6 Max · proprietary | 50.0% | API only |
| 32 | DeepSeek V4 Pro 0813 · 1650.5B | 50.0% | ~991 GB |
| 33 | GPT 5.5 (xhigh) · proprietary | 50.0% | API only |
| 34 | GLM 5.3 · 753.3B | 49.0% | ~456 GB |
| 35 | Gemini 3.8 Flash (unspecified) · proprietary | 48.0% | API only |
| 36 | Claude Sonnet 4.6 Max · proprietary | 45.0% | API only |
| 37 | Muse Spark 1.2 · proprietary | 43.0% | API only |
| 38 | Step 5 Preview (unspecified) · proprietary | 42.0% | API only |
| 39 | Muse Spark 1.1 · proprietary | 39.0% | API only |
| 40 | Claude Opus 4.5 (Nov 01, 2025, unspecified) · proprietary | 36.0% | API only |
| 41 | Gemini 3.6 Flash (unspecified) · proprietary | 36.0% | API only |
| 42 | GLM 5.2 · 753.3B | 35.0% | ~456 GB |
| 43 | Grok 4.7 (unspecified) · proprietary | 34.0% | API only |
| 44 | Gemini 3.5 Flash (high) · proprietary | 31.0% | API only |
| 45 | Grok 4.5 (high) · proprietary | 31.0% | API only |
| 46 | Gemini 3.1 Pro Preview · proprietary | 26.0% | API only |
| 47 | Qwen3.7 Max · proprietary | 26.0% | API only |
| 48 | GLM 5.1 · 753.9B | 22.2% | ~457 GB |
| 49 | MiMo V2.5 Pro · 1023.2B | 22.0% | ~614 GB |
| 50 | GLM 5.3 Flash · 321.3B | 21.0% | ~195 GB |
| 51 | GPT 5.4 Mini (Mar 17, 2026, xhigh) · proprietary | 21.0% | API only |
| 52 | Gemini 3 Pro Preview · proprietary | 20.0% | API only |
| 53 | Claude Sonnet 4.5 (Sep 29, 2025, unspecified) · proprietary | 19.0% | API only |
| 54 | GPT 5 (Aug 07, 2025, high) · proprietary | 18.0% | API only |
| 55 | MiniMax M3 · 427.0B | 18.0% | ~257 GB |
| 56 | Muse Spark · proprietary | 17.0% | API only |
| 57 | DeepSeek V4 Pro · 1598.8B | 16.0% | ~960 GB |
| 58 | Kimi K2.6 · 1026.9B | 16.0% | ~620 GB |
| 59 | MiMo V2.5 · 310.8B | 16.0% | ~187 GB |
| 60 | Qwen3.8 27B · 27.8B | 16.0% | ~17 GB |
| 61 | Gemini 3 Flash Preview · proprietary | 15.0% | API only |
| 62 | GPT 5.2 (Dec 11, 2025, xhigh) · proprietary | 15.0% | API only |
| 63 | Grok 4.20 0309 Reasoning · proprietary | 14.0% | API only |
| 64 | Gemini 3.5 Flash Lite · proprietary | 13.0% | API only |
| 65 | GPT 5 Nano (Aug 07, 2025, high) · proprietary | 12.0% | API only |
| 66 | Grok 4.3 (high) · proprietary | 11.0% | API only |
| 67 | Mistral Large 4.0 (unspecified) · proprietary | 10.0% | API only |
| 68 | GPT 5 Mini (Aug 07, 2025, high) · proprietary | 9.0% | API only |
| 69 | GPT 5.1 Codex Max · proprietary | 9.0% | API only |
| 70 | Mistral Medium 2604 · proprietary | 9.0% | API only |
| 71 | DeepSeek V3.2 · 685.4B | 8.0% | ~415 GB |
| 72 | GLM 4.7 · 358.3B | 6.0% | ~216 GB |
| 73 | Inkling Small · 266.0B | 6.0% | ~160 GB |
| 74 | GPT 5.4 Nano (Mar 17, 2026, high) · proprietary | 5.0% | API only |
| 75 | Grok 4.1 Fast Reasoning · proprietary | 4.0% | API only |
| 76 | MiniMax M2.5 · 228.7B | 4.0% | ~138 GB |
| 77 | Mercury 2.5 (unspecified) · proprietary | 3.0% | API only |
| 78 | MiniMax M2.7 · 228.7B | 3.0% | ~138 GB |
| 79 | NVIDIA Nemotron 3 Ultra 550B A55B BF16 · 560.5B | 2.0% | ~370 GB |
| 80 | Inkling · 952.4B | 0.0% | ~572 GB |
| 81 | Laguna M.1 · 225.8B | 0.0% | ~136 GB |
| 82 | Laguna XS.2 · 33.4B | 0.0% | ~20 GB |
Score vs model size
Which models give the most quality for their size — the ones worth running locally.
- Qwen3.8 27B, 28B, score 16.0% — on the efficiency frontier (best score at its size or smaller).
- DeepSeek V4 Flash 0731, 304B, score 56.0% — on the efficiency frontier (best score at its size or smaller).
- MiMo V2.6 Flash RL, 309B, score 63.0% — on the efficiency frontier (best score at its size or smaller).
- Hy4 Preview, 780B, score 75.0% — on the efficiency frontier (best score at its size or smaller).
- Kimi K3, 2.8T, score 87.0% — on the efficiency frontier (best score at its size or smaller).
ProofBench: frequently asked questions
- What is the best open LLM on ProofBench?
- Kimi K3 is the top open model on ProofBench, scoring 87.0%. Among all models tested — including proprietary ones — it ranks #10. The top model overall is Claude Fable 5.1 Max (Anthropic) at 100.0%.
- What's the best ProofBench model you can run on a 24 GB GPU?
- Qwen3.8 27B is the highest-scoring open model that fits in 24 GB at 4-bit quantization (~17 GB), scoring 16.0% on ProofBench.
- Can open models match proprietary models on ProofBench?
- Not quite on ProofBench: the strongest proprietary model (Claude Fable 5.1 Max) scores 100.0%, ahead of the best open model (Kimi K3) at 87.0% — but you can run the open one yourself.
Scores aggregated from epoch. llmrun does not run this benchmark — see the source for methodology, or the about benchmarks for what it measures.