Humanity's Last Exam Diamond
General QA
Humanity's Last Exam Diamond is a fixed 1,000-question revision created by Scale AI and the Center for AI Safety, with 500 reasoning and 500 knowledge questions drawn from HLE and its held-out reserve. The leaderboard reports overall accuracy under closed-book evaluation with automatic answer extraction and judging.
Top models (higher is better)
| Model | Score |
|---|---|
| Gemini 4 Carbon | 69.1 |
| GPT-6 Astra | 63.8 |
| GPT-6.1 Sol | 58.8 |
| Claude Opus 5.5 | 55.0 |
| Claude Sonnet 5.5 | 53.3 |
| Gemini 4 Argon | 53.2 |
| Claude Fable 5.1 | 51.3 |
| Muse Wonder Full | 48.5 |
| GPT-6.1 Terra | 47.3 |
| Claude Fable 5 | 45.6 |
| GPT-6 Sol | 45.2 |
| Kimi K3.1 Pro | 39.4 |
| Muse Wonder | 39.0 |
| Claude Opus 5 | 38.6 |
| Gemini 3.8 Flash | 34.3 |
| Muse Spark 1.3 | 33.7 |
| Gemini 3.1 Pro Preview | 31.8 |
| Muse Spark 1.1 | 31.3 |
| GPT-5.6 Sol | 31.2 |
| Gemini 3.7 Flash | 31.2 |
| Claude Haiku 5.5 | 29.6 |
| GPT-5.6 Terra | 28.8 |
| MiniMax M3.1 Flash | 28.1 |
| Muse Spark Open | 28.0 |
| GPT-5.5 | 27.9 |
| Kimi K3 | 26.9 |
| Gemini 3.5 Flash | 25.7 |
| DeepSeek V4.1 Flash 1017 | 25.3 |
| Step 5 Preview | 24.3 |
| Qwen3.8 Max (0902) | 24.0 |
| Grok 4.7 | 23.6 |
| Opus 4.8 | 23.1 |
| GLM-5.2 | 22.3 |
| GPT-6 Luna | 22.2 |
| DeepSeek V4 Flash Vision Exp | 21.1 |
| GPT-5.4 | 21.1 |
| Opus 4.7 | 20.4 |
| Gemini 3 Pro Preview | 20.2 |
| Grok 4.6 | 19.6 |
| DeepSeek V4 Pro 0813 | 19.5 |