Math
FrontierMath Leaderboard
FrontierMath is a benchmark of exceptionally hard, original research-level mathematics problems created with professional mathematicians. Even the strongest models solve only a small fraction, making it a frontier measure of genuine mathematical ability.
Source: epoch17 open models ranked+97 proprietaryData through Oct 2026
All models ranked on FrontierMath
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 | GPT 6 Astra Max · proprietary | 93.7% | API only |
| 2 | GPT 6.1 Sol Max · proprietary | 93.7% | API only |
| 3 | Claude Opus 5.5 Max · proprietary | 91.2% | API only |
| 4 | Claude Fable 5.1 Max · proprietary | 90.2% | API only |
| 5 | GPT 6 Sol Max · proprietary | 89.8% | API only |
| 6 | GPT 5.6 Sol Max · proprietary | 89.1% | API only |
| 7 | Claude Sonnet 5.5 Max · proprietary | 88.8% | API only |
| 8 | GPT 5.5 Pro (xhigh) · proprietary | 87.7% | API only |
| 9 | Claude Fable 5 Max · proprietary | 87.0% | API only |
| 10 | Gemini 4 Argon (high) · proprietary | 86.0% | API only |
| 11 | GPT 5.6 Terra Max · proprietary | 86.0% | API only |
| 12 | Claude Opus 5 Max · proprietary | 85.6% | API only |
| 13 | GPT 5.5 (xhigh) · proprietary | 85.3% | API only |
| 14 | GPT 5.4 Pro (Mar 05, 2026, xhigh) · proprietary | 82.5% | API only |
| 15 | GPT 5.6 Luna Max · proprietary | 82.1% | API only |
| 16 | Muse Wonder 1.0 Max · proprietary | 82.1% | API only |
| 17 | Claude Opus 4.8 Max · proprietary | 80.0% | API only |
| 18 | GPT 6 Luna Max · proprietary | 79.0% | API only |
| 19 | GPT 5.4 (Mar 05, 2026, xhigh) · proprietary | 78.6% | API only |
| 20 | Claude Haiku 5.5 Max · proprietary | 75.1% | API only |
| 21 | Qwen3.8 Max (xhigh) · proprietary | 74.7% | API only |
| 22 | Muse Spark 1.3 (xhigh) · proprietary | 74.4% | API only |
| 23 | Muse Spark 1.3 Max · proprietary | 74.0% | API only |
| 24 | GPT 5.2 Pro (Dec 11, 2025, xhigh) · proprietary | 74.0% | API only |
| 25 | Kimi K3 · 2779.9B | 72.2% | ~1674 GB |
| 26 | Gemini 3.7 Flash (high) · proprietary | 71.6% | API only |
| 27 | Claude Opus 4.7 Max · proprietary | 70.2% | API only |
| 28 | Muse Spark Open 816B A42B · 822.0B | 69.8% | ~543 GB |
| 29 | GLM 5.3 · 753.3B | 68.8% | ~456 GB |
| 30 | Gemini 3.8 Flash (high) · proprietary | 68.4% | API only |
| 31 | GPT 5.2 (Dec 11, 2025, xhigh) · proprietary | 67.4% | API only |
| 32 | DeepSeek V4.1 Flash · 763.2B | 67.4% | ~458 GB |
| 33 | Claude Opus 4.6 Max · proprietary | 66.0% | API only |
| 34 | Grok 4.6 (xhigh) · proprietary | 66.0% | API only |
| 35 | Claude Sonnet 5 Max · proprietary | 65.6% | API only |
| 36 | Qwen3.8 Max (Sep 02, xhigh) · proprietary | 65.6% | API only |
| 37 | DeepSeek V4 Pro 0813 · 1650.5B | 64.6% | ~991 GB |
| 38 | Qwen3.7 Max · proprietary | 64.6% | API only |
| 39 | Gemini 3.5 Flash (high) · proprietary | 62.8% | API only |
| 40 | Gemini 3.1 Pro Preview · proprietary | 59.7% | API only |
| 41 | GLM 5.2 · 753.3B | 59.2% | ~456 GB |
| 42 | Gemini 3.6 Flash (high) · proprietary | 59.0% | API only |
| 43 | Mistral Large 4.0 (high) · proprietary | 58.6% | API only |
| 44 | DeepSeek V4 Flash 0731 · 304.2B | 57.5% | ~183 GB |
| 45 | Grok 4.5 (high) · proprietary | 57.2% | API only |
| 46 | Kimi K2.6 · 1026.9B | 57.2% | ~620 GB |
| 47 | GLM 5.3 Flash · 321.3B | 55.8% | ~195 GB |
| 48 | GPT 5 Pro (Oct 06, 2025, high) · proprietary | 55.8% | API only |
| 49 | GPT 5 (Aug 07, 2025, high) · proprietary | 55.4% | API only |
| 50 | Kimi K2.7 Code · 1026.9B | 54.0% | ~620 GB |
| 51 | Grok 4.7 (xhigh) · proprietary | 53.0% | API only |
| 52 | Gemini 3 Flash Preview · proprietary | 51.2% | API only |
| 53 | GPT 5.4 Mini (Mar 17, 2026, xhigh) · proprietary | 51.2% | API only |
| 54 | GPT 5 Mini (Aug 07, 2025, high) · proprietary | 46.7% | API only |
| 55 | Inkling Small · 266.0B | 46.3% | ~160 GB |
| 56 | DeepSeek V4 Pro · 1598.8B | 45.3% | ~960 GB |
| 57 | GPT 5.4 Nano (Mar 17, 2026, high) · proprietary | 44.9% | API only |
| 58 | Grok 4.20 0309 Reasoning · proprietary | 44.9% | API only |
| 59 | Grok 4.3 (high) · proprietary | 42.8% | API only |
| 60 | GPT 5.6 Luna (low) · proprietary | 41.4% | API only |
| 61 | GPT 5.6 Luna None · proprietary | 39.6% | API only |
| 62 | Qwen3.6 Plus · proprietary | 38.3% | API only |
| 63 | GPT 5 (Aug 07, 2025, low) · proprietary | 37.2% | API only |
| 64 | GLM 5.1 · 753.9B | 36.8% | ~457 GB |
| 65 | O4 Mini (Apr 16, 2025, high) · proprietary | 36.1% | API only |
| 66 | Qwen3.6 27B · 27.8B | 35.1% | ~17 GB |
| 67 | Claude Opus 4.5 (Nov 01, 2025, 32K) · proprietary | 34.4% | API only |
| 68 | Qwen3.7 Plus None · proprietary | 34.4% | API only |
| 69 | Inkling · 952.4B | 33.3% | ~572 GB |
| 70 | O3 (Apr 16, 2025, high) · proprietary | 33.3% | API only |
| 71 | Qwen3.6 Plus None · proprietary | 32.3% | API only |
| 72 | Qwen3.5 397B A17B · 403.4B | 31.2% | ~242 GB |
| 73 | O3 (Apr 16, 2025, medium) · proprietary | 29.8% | API only |
| 74 | O4 Mini (Apr 16, 2025, medium) · proprietary | 28.8% | API only |
| 75 | Gemini 3.1 Flash Lite (high) · proprietary | 27.7% | API only |
| 76 | GPT 5.5 Instant · proprietary | 26.3% | API only |
| 77 | Gemini 3.5 Flash Lite (high) · proprietary | 26.0% | API only |
| 78 | Gemini 2.5 Pro · proprietary | 24.6% | API only |
| 79 | GPT 5.4 Mini (Mar 17, 2026, low) · proprietary | 24.6% | API only |
| 80 | Claude Sonnet 4.5 (Sep 29, 2025, 32K) · proprietary | 23.9% | API only |
| 81 | Gemini 3.1 Flash Lite (low) · proprietary | 22.5% | API only |
| 82 | Qwen3.6 Flash None · proprietary | 22.5% | API only |
| 83 | Gemini 3.1 Flash Lite (minimal) · proprietary | 21.4% | API only |
| 84 | GPT 5.4 Nano (Mar 17, 2026, low) · proprietary | 20.3% | API only |
| 85 | Qwen3.6 35B A3B · 36.0B | 20.3% | ~22 GB |
| 86 | GPT 5 Nano (Aug 07, 2025, high) · proprietary | 20.0% | API only |
| 87 | O3 (Apr 16, 2025, low) · proprietary | 19.3% | API only |
| 88 | Qwen3.7 Flash · proprietary | 19.3% | API only |
| 89 | Qwen3.7 Flash None · proprietary | 19.3% | API only |
| 90 | Qwen3 Max (Sep 23, 2025) · proprietary | 18.9% | API only |
| 91 | O3 Mini (Jan 31, 2025, high) · proprietary | 18.6% | API only |
| 92 | GPT 5 (Aug 07, 2025, minimal) · proprietary | 18.3% | API only |
| 93 | GPT 5 Mini (Aug 07, 2025, low) · proprietary | 18.3% | API only |
| 94 | Qwen3.5 Flash None · proprietary | 18.3% | API only |
| 95 | GPT 5.4 Mini 2026 03.17 None · proprietary | 17.2% | API only |
| 96 | Qwen3.6 Flash · proprietary | 17.2% | API only |
| 97 | O4 Mini (Apr 16, 2025, low) · proprietary | 16.1% | API only |
| 98 | O1 (Dec 17, 2024, high) · proprietary | 14.7% | API only |
| 99 | Claude Opus 4.1 (Aug 05, 2025, 32K) · proprietary | 12.6% | API only |
| 100 | O3 Mini (Jan 31, 2025, medium) · proprietary | 10.5% | API only |
| 101 | O1 (Dec 17, 2024, medium) · proprietary | 10.2% | API only |
| 102 | Qwen3.5 Flash · proprietary | 9.5% | API only |
| 103 | O1 (Dec 17, 2024, low) · proprietary | 8.4% | API only |
| 104 | GPT 4.1 Mini (Apr 14, 2025) · proprietary | 6.7% | API only |
| 105 | GPT 4.1 (Apr 14, 2025) · proprietary | 6.0% | API only |
| 106 | GPT 5 Mini (Aug 07, 2025, minimal) · proprietary | 6.0% | API only |
| 107 | GPT 5 Nano (Aug 07, 2025, low) · proprietary | 6.0% | API only |
| 108 | GPT 5.4 Nano 2026 03.17 None · proprietary | 4.6% | API only |
| 109 | O3 Mini (Jan 31, 2025, low) · proprietary | 3.9% | API only |
| 110 | GPT 5 Nano (Aug 07, 2025, minimal) · proprietary | 1.8% | API only |
| 111 | GPT 4 Turbo (Apr 09, 2024) · proprietary | 0.7% | API only |
| 112 | GPT 4o Mini (Jul 18, 2024) · proprietary | 0.7% | API only |
| 113 | GPT 4o (Aug 06, 2024) · proprietary | 0.4% | API only |
| 114 | GPT 3.5 Turbo (Jan 25) · proprietary | 0.0% | API only |
Score vs model size
Which models give the most quality for their size — the ones worth running locally.
- Qwen3.6 27B, 28B, score 35.1% — on the efficiency frontier (best score at its size or smaller).
- Inkling Small, 266B, score 46.3% — on the efficiency frontier (best score at its size or smaller).
- DeepSeek V4 Flash 0731, 304B, score 57.5% — on the efficiency frontier (best score at its size or smaller).
- GLM 5.3, 753B, score 68.8% — on the efficiency frontier (best score at its size or smaller).
- Muse Spark Open 816B A42B, 822B, score 69.8% — on the efficiency frontier (best score at its size or smaller).
- Kimi K3, 2.8T, score 72.2% — on the efficiency frontier (best score at its size or smaller).
FrontierMath: frequently asked questions
- What is the best open LLM on FrontierMath?
- Kimi K3 is the top open model on FrontierMath, scoring 72.2%. Among all models tested — including proprietary ones — it ranks #25. The top model overall is GPT 6.1 Sol Max (OpenAI) at 93.7%.
- What's the best FrontierMath model you can run on a 24 GB GPU?
- Qwen3.6 27B is the highest-scoring open model that fits in 24 GB at 4-bit quantization (~17 GB), scoring 35.1% on FrontierMath.
- Can open models match proprietary models on FrontierMath?
- Not quite on FrontierMath: the strongest proprietary model (GPT 6.1 Sol Max) scores 93.7%, ahead of the best open model (Kimi K3) at 72.2% — 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.