MMLU Pro
Broad academic knowledge across fourteen subjects with ten answer options instead of four.
KnowledgeTIGER-Lab138 models run by Vals AIUpdated 1 September 2026
How it is run
A harder replacement for the original MMLU, designed to reduce the effect of guessing and of test-set contamination.
What it cannot tell you
Almost every frontier model now scores between 86% and 92%. This benchmark no longer separates frontier models; it separates them from 2025 models.
- Leader
- Claude Fable 5.192.38% · Vals AI's own list
- Top-ten spread
- 2.86%Leader minus tenth place, or last place in smaller fields.
- Median
- 88.44%of the 24 models on this site
- No published result
- 22of 46 models on this site
Independent leaderboard · Vals AI
Top 138 of 138 runs by Vals AI, one harness.
| # | Model | Score |
|---|---|---|
| 1 | Claude Fable 5.1 | 92.38% |
| 2 | Claude Opus 5 | 91.59% |
| 3 | Claude Fable 5 | 91.50% |
| 4 | Gemini 3.1 Pro | 90.99% |
| 5 | Gemini 3.8 Flash | 90.22% |
| 6 | Gemini 3.7 Flash | 90.12% |
| 7 | Gemini 3 Pro Preview | 90.10% |
| 8 | Claude Opus 4.7 | 89.87% |
| 9 | Claude Opus 4.8 | 89.59% |
| 10 | Gemini 3.5 Flash | 89.52% |
| 11 | Grok 4.6 | 89.40% |
| 12 | Qwen3.7 Max | 89.31% |
| 13 | Gemini 3.6 Flash | 89.28% |
| 14 | Grok 4.5 | 89.22% |
| 15 | Claude Opus 4 6 Thinking | 89.11% |
| 16 | GPT-5.6 Sol | 89.10% |
| 17 | Muse Spark 1 1 | 88.73% |
| 18 | Qwen3.8 Max | 88.60% |
| 19 | Gemini 3 Flash Preview | 88.59% |
| 20 | Muse Spark 1.2 | 88.28% |
| 21 | GPT-5.5 | 88.14% |
| 22 | Kimi K3 | 87.97% |
| 23 | Claude Opus 4 1 20250805 Thinking | 87.92% |
| 24 | Qwen3.6 Plus | 87.67% |
| 25 | Kimi K2.6 | 87.57% |
| 26 | Claude Sonnet 5 | 87.55% |
| 27 | Gpt 5.4 2026 03 05 | 87.48% |
| 28 | Claude Sonnet 4 5 20250929 Thinking | 87.36% |
| 29 | Claude Sonnet 4 6 | 87.34% |
| 30 | Muse Spark | 87.32% |
| 31 | Claude Opus 4 5 20251101 Thinking | 87.26% |
| 32 | Deepseek V4 Pro | 87.25% |
| 33 | Claude Opus 4 1 20250805 | 87.21% |
| 34 | Qwen3.5 Plus Thinking | 87.18% |
| 35 | MiniMax M2.1 | 87.05% |
| 36 | DeepSeek V4 Pro | 86.97% |
| 37 | Glm 5.1 | 86.90% |
| 38 | GLM 5.3 | 86.77% |
| 39 | Glm 5.2 | 86.71% |
| 40 | GPT-5.6 Terra | 86.66% |
| 41 | Gpt 5 2025 08 07 | 86.54% |
| 42 | Gpt 5.1 2025 11 13 | 86.38% |
| 43 | Inkling | 86.30% |
| 44 | Grok 4.20 0309 Reasoning | 86.25% |
| 45 | Gemini 3.1 Flash Lite Preview | 86.24% |
| 46 | Gpt 5.2 2025 12 11 | 86.23% |
| 47 | Deepseek V4 Flash 0731 | 86.21% |
| 48 | Claude Opus 4 20250514 | 86.17% |
| 49 | GLM 5.3 Flash | 86.06% |
| 50 | GPT-5.6 Luna | 86.04% |
| 51 | Glm 5 Thinking | 86.03% |
| 52 | Kimi K2.5 Thinking | 85.91% |
| 53 | Grok 4.3 | 85.84% |
| 53 | Gemini 3.5 Flash Lite | 85.84% |
| 55 | Nemotron 3 Ultra 550b A55b | 85.76% |
| 56 | O3 2025 04 16 | 85.60% |
| 57 | Claude Opus 4 5 20251101 | 85.59% |
| 58 | Inkling Small | 85.57% |
| 59 | Grok 4 0709 | 85.30% |
| 60 | Qwen3 Max 2026 01 23 | 84.98% |
| 61 | Deepseek V3p2 Thinking | 84.92% |
| 62 | Mimo V2.5 Pro | 84.60% |
| 63 | Gpt 5.4 Mini 2026 03 17 | 84.55% |
| 64 | Qwen3 Max | 84.36% |
| 65 | Qwen3.8 27b | 84.34% |
| 66 | MiniMax M3 | 84.22% |
| 67 | Grok 4 1 Fast Reasoning | 84.18% |
| 68 | Qwen3.5 Flash | 84.06% |
| 68 | Gemini 2.5 Pro Exp 03 25 | 84.06% |
| 70 | Claude Sonnet 4 20250514 Thinking | 83.87% |
| 71 | Gemini 2.5 Flash Preview 09 2025 | 83.69% |
| 72 | Gemini 2.5 Flash Preview 09 2025 Thinking | 83.66% |
| 73 | Qwen3 Max Preview | 83.54% |
| 74 | O1 2024 12 17 | 83.49% |
| 75 | Deepseek R1 | 83.18% |
| 76 | Deepseek V3p2 | 83.06% |
| 77 | Mimo V2.5 | 82.93% |
| 78 | Glm 4.7 | 82.74% |
| 79 | Claude 3 7 Sonnet 20250219 Thinking | 82.73% |
| 80 | Gpt 5 Mini 2025 08 07 | 82.23% |
| 81 | Glm 4.6 | 82.20% |
| 82 | Ling 3.0 Flash 2607 | 82.01% |
| 83 | Grok 3 Mini Fast High Reasoning | 81.37% |
| 84 | Qwen3 235b A22b | 81.25% |
| 85 | Glm 4.5 | 81.22% |
| 86 | Kimi K2 Thinking | 81.07% |
| 87 | Claude 3 7 Sonnet 20250219 | 80.66% |
| 88 | O4 Mini 2025 04 16 | 80.56% |
| 89 | Gpt 4.1 2025 04 14 | 80.50% |
| 90 | MiniMax M2.7 | 80.43% |
| 91 | MiniMax M2.5 | 80.09% |
| 92 | Grok 3 Mini Fast Low Reasoning | 80.01% |
| 93 | Grok 3 | 79.95% |
| 94 | Mistral Large 2512 | 79.82% |
| 95 | Grok 4 Fast Reasoning | 79.70% |
| 96 | Deepseek V3 0324 | 79.47% |
| 97 | Claude Sonnet 4 20250514 | 79.43% |
| 98 | Llama4 Maverick Instruct Basic | 79.42% |
| 99 | Kimi K2 Instruct | 79.39% |
| 100 | Gpt Oss 120b | 79.17% |
| 101 | Gemini 2.5 Flash Lite Preview 09 2025 Thinking | 79.12% |
| 102 | Claude Haiku 4 5 20251001 Thinking | 78.72% |
| 103 | O3 Mini 2025 01 31 | 78.69% |
| 104 | Gemini 2.5 Flash Lite Preview 09 2025 | 78.64% |
| 105 | Claude 3 5 Sonnet 20241022 | 78.40% |
| 106 | Gemini 2.0 Flash 001 | 77.38% |
| 107 | Gpt 4.1 Mini 2025 04 14 | 77.22% |
| 108 | Gpt 5.4 Nano 2026 03 17 | 77.17% |
| 109 | Gpt 5 Nano 2025 08 07 | 76.07% |
| 110 | Grok 2 1212 | 75.47% |
| 111 | Mistral Medium 3.5 | 75.33% |
| 112 | Gemini 1.5 Pro 002 | 75.29% |
| 112 | Mistral Medium 2505 | 75.29% |
| 114 | Grok 4 1 Fast Non Reasoning | 75.21% |
| 115 | Gpt 4o 2024 08 06 | 74.13% |
| 116 | Deepseek V3 | 73.82% |
| 117 | Gpt 4o 2024 11 20 | 72.56% |
| 118 | Gpt Oss 20b | 71.64% |
| 119 | Llama 3.3 Nemotron Super 49b V1 42e84561 | 70.78% |
| 120 | Grok 4 Fast Non Reasoning | 70.34% |
| 121 | Llama 3.3 70B Instruct Turbo | 69.86% |
| 122 | Mistral Large 2411 | 69.71% |
| 123 | Llama 4 Scout 17B 16E Instruct | 69.63% |
| 124 | Llama 3.3 Nemotron Super 49b V1 42e84561 Thinking | 69.58% |
| 125 | Command A 03 2025 | 69.17% |
| 126 | Laguna Xs.2 | 69.05% |
| 127 | Laguna M.1 | 68.84% |
| 128 | Magistral Medium 2509 | 68.66% |
| 129 | Mistral Small 2503 | 66.02% |
| 130 | Gemini 1.5 Flash 002 | 65.61% |
| 131 | Mistral Small 2402 | 64.44% |
| 132 | Claude 3 5 Haiku 20241022 | 64.12% |
| 133 | Gpt 4.1 Nano 2025 04 14 | 63.48% |
| 134 | Gpt 4o Mini 2024 07 18 | 62.74% |
| 135 | Magistral Small 2509 | 62.14% |
| 136 | Jamba Large 1.6 | 49.78% |
| 137 | Command R Plus | 44.00% |
| 138 | Jamba Mini 1.6 | 30.28% |
No published result tracked
22 of the 46 models on this site have no published number on this benchmark, so their cells in the home table stay blank rather than reading as zero.
Step 5 PreviewGPT-6 LunaQwen3.8 2.4T A95BQwen3.8-Flash-NextMiMo-V2.6-FlashQwen3.8 Max (0902)GPT-5.4DeepSeek V4.1 Flash 1017GLM-LegalClaude Haiku 5.5BeamGemini 4 ArgonClaude Sonnet 5.5Claude Opus 5.5GPT-6 AstraGPT-6 SolGPT-6.1 SolGrok 4.7DeepSeek V4.1 FlashMuse Spark 1.3 MaxHunyuan Hy4 PreviewMiMo V2.6 Pro
Sources
Every page used for this benchmark, read on 17 October 2026.
- Benchmark site or the page that publishes ithuggingface.co/datasets/TIGER-Lab/MMLU-Pro
- Independent resultswww.vals.ai/benchmarks/mmlu_pro