Humanity's Last Exam Diamond (Reasoning Partition)
General QA
The 500-question reasoning partition of Humanity's Last Exam Diamond (the Scale AI and Center for AI Safety 1,000-question revision, split evenly between reasoning and knowledge questions). Accuracy is the share of the partition's questions judged correct, pass@1, from the same run as the overall HLE-Diamond score.
Top models (higher is better)
| Model | Score |
|---|---|
| Gemini 4 Carbon | 79.6 |
| GPT-6 Astra | 78.0 |
| GPT-6.1 Sol | 76.2 |
| Claude Sonnet 5.5 | 68.0 |
| GPT-6.1 Terra | 65.4 |
| Gemini 4 Argon | 64.8 |
| Muse Wonder Full | 61.2 |
| GPT-6 Sol | 59.6 |
| Muse Wonder | 50.4 |
| Claude Haiku 5.5 | 44.8 |
| Kimi K3.1 Pro | 44.8 |
| Muse Spark 1.3 | 43.0 |
| GPT-5.6 Terra | 40.4 |
| Gemini 3.8 Flash | 37.2 |
| DeepSeek V4.1 Flash 1017 | 34.8 |
| Gemini 3.1 Pro Preview | 34.8 |
| MiniMax M3.1 Flash | 34.6 |
| Qwen3.8 Max (0902) | 33.6 |
| Kimi K3 | 33.0 |
| Muse Spark Open | 33.0 |
| Grok 4.7 | 31.2 |
| GPT-6 Luna | 31.0 |
| DeepSeek V4 Flash Vision Exp | 27.2 |
| Step 5 Preview | 26.6 |
| DeepSeek V4 Pro 0813 | 25.2 |
| DeepSeek V4.1 Flash | 24.6 |
| Mistral Large 4 | 23.6 |
| GLM-5.3 | 23.4 |
| Qwen3.8-Flash-Next | 23.4 |
| Ling 3.1 Flash 1019 | 23.2 |
| GLM-5.3 Flash | 22.0 |
| Mistral Large 4 Preview | 22.0 |
| Qwen3.8 27B | 22.0 |
| MiMo-V2.6-Pro | 19.0 |
| Quasar 438B | 18.8 |
| Kimi K2.7 Code | 18.4 |
| Qwen3.7-Plus (source unspecified) | 18.0 |
| Beam 1.0 501B A23B | 17.8 |
| GPT-5.5 Instant (2026-06-25 hosted snapshot) | 17.8 |
| MiMo-V2.6-Flash | 17.4 |