Humanity's Last Exam (HLE) humanitys-last-exam Leaderboard
Humanity's Last Exam — 2,500 expert-authored frontier questions across 100+ subjects (text + multimodal). The premier "is the frontier still hard" benchmark; top models score well under half. Accuracy, higher is better. · Metric: Accuracy (higher is better) · 🟢 Updated 2h ago
| # | Model | Accuracy | Paper |
|---|---|---|---|
| 1 | Claude Fable 5 (Adaptive Reasoning, Max Effort, Opus 4.8 Fallback) | 53.30 | — |
| 2 | GPT-5.6 Sol (max) | 47.20 | — |
| 3 | Claude Opus 4.8 (Adaptive Reasoning, Max Effort) | 45.70 | — |
| 4 | Gemini 3.1 Pro Preview | 44.70 | — |
| 5 | GPT-5.6 Sol (xhigh) | 44.70 | — |
| 6 | GPT-5.5 (xhigh) | 44.30 | — |
| 7 | GPT-5.6 Sol (high) | 44.10 | — |
| 8 | GPT-5.5 (high) | 43.00 | — |
| 9 | GPT-5.6 Terra (max) | 41.80 | — |
| 10 | GPT-5.4 (xhigh) | 41.60 | — |
| 11 | Gemini 3.5 Flash (high) | 41.00 | — |
| 12 | GPT-5.5 (medium) | 40.60 | — |
| 13 | Grok 4.5 (high) | 40.30 | — |
| 14 | GLM-5.2 (max) | 40.10 | — |
| 15 | GPT-5.6 Terra (xhigh) | 40.00 | — |
| 16 | Gemini 3.5 Flash (medium) | 39.90 | — |
| 17 | GPT-5.3 Codex (xhigh) | 39.90 | — |
| 18 | Muse Spark | 39.90 | — |
| 19 | GPT-5.6 Sol (medium) | 39.70 | — |
| 20 | Claude Opus 4.7 (Adaptive Reasoning, Max Effort) | 39.60 | — |
| 21 | Claude Sonnet 5 (Adaptive Reasoning, Max Effort) | 39.60 | — |
| 22 | Qwen3.7 Max | 38.10 | — |
| 23 | Gemini 3 Pro Preview (high) | 37.20 | — |
| 24 | GPT-5.6 Luna (max) | 37.20 | — |
| 25 | MiniMax-M3 | 37.10 | — |
| 26 | Claude Opus 4.6 (Adaptive Reasoning, Max Effort) | 36.70 | — |
| 27 | GPT-5.6 Terra (high) | 36.70 | — |
| 28 | GPT-5.6 Sol (low) | 36.60 | — |
| 29 | Grok Build 0.1 0616 | 36.00 | — |
| 30 | DeepSeek V4 Pro (Reasoning, Max Effort) | 35.90 | — |