Humanity's Last Exam: leaderboard
PhD-level questions at the frontier of human expert knowledge. Made by the same team behind MASK. Even top models show high calibration errors, indicating reasoning gaps.
Metric: Score. Source: scale.com. Status: years away from saturation. 51 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Claude Fable 5.1 (xHigh) | 46.5 |
| 2 | Gemini 3.1 Pro (Preview) (High) | 46.44 |
| 3 | GPT-5.4 Pro (xHigh) | 44.32 |
| 4 | Muse Spark | 40.56 |
| 5 | Gemini 3 Pro (Preview) | 37.52 |
| 6 | GPT-5.4 (xHigh) | 36.24 |
| 7 | Claude Opus 4.7 | 36.2 |
| 8 | Claude Opus 4.6 (Adaptive Reasoning, Max Effort) | 34.44 |
| 9 | GPT-5 Pro | 31.64 |
| 10 | GPT-5.2 | 27.8 |
| 11 | GPT-5 | 25.32 |
| 12 | Claude Opus 4.5 (20251101) (Thinking) | 25.2 |
| 13 | Kimi K2.5 | 24.37 |
| 14 | GPT-5.1 (Thinking) | 23.68 |
| 15 | Gemini 2.5 Pro (Preview 06-05) | 21.64 |
Interactive version: theaggregate.ai/benchmark?slug=humanity-s-last-exam · How It Works · Data refreshed daily, snapshot 2026-09-05.