Reasoning
Humanitys-Last-Exam
The hardest known benchmark — novel academic problems.
20 models published a score
| # | Model | Company | Score |
|---|---|---|---|
| 1 | Muse Spark | Meta | 58.0 |
| 2 | Claude Mythos Preview | Anthropic | 56.8 |
| 3 | Claude Opus 4.7 | Anthropic | 54.7 |
| 4 | Grok 4 Heavy | xAI | 50.7 |
| 5 | GLM-5 | Zhipu AI | 50.4 |
| 6 | Kimi K2.5 | Moonshot AI | 50.2 |
| 7 | Gemini 3 Deep Think | Google DeepMind | 48.4 |
| 8 | Gemini 3.1 Pro | Google DeepMind | 44.4 |
| 9 | GLM-4.7 | Zhipu AI | 42.8 |
| 10 | Claude Opus 4.6 | Anthropic | 40.0 |
| 11 | DeepSeek V4 Pro | DeepSeek | 37.7 |
| 12 | Gemini 3 Pro | Google DeepMind | 37.5 |
| 13 | Kimi K2.6 | Moonshot AI | 34.7 |
| 14 | Gemini 3 Flash | Google DeepMind | 33.7 |
| 15 | MiMo V2 Pro | Xiaomi | 28.3 |
| 16 | Grok 4 | xAI | 25.4 |
| 17 | MiMo V2 Flash | Xiaomi | 20.0 |
| 18 | Nemotron 3 Super | Nvidia | 17.4 |
| 19 | Gemini 3.1 Flash-Lite | Google DeepMind | 16.0 |
| 20 | K-EXAONE 236B-A23B | LG AI Research | 13.6 |