Humanity's Last Exam as reported in the LLM Benchmark Matrix. Crowd-sourced from experts across many fields to be the hardest possible questions for AI.
This definition has not completed editorial and provenance review. Verify its metric, version, source, and evaluation configuration upstream before using it for a model-selection decision.
Extremely high difficulty ceiling — designed to remain challenging as models improve. Broad domain coverage.
Matrix snapshot — we also have AA's own HLE column. Very hard: even top models score under 50%.
Higher is better. Accuracy percentage (0–100%). Top frontier models: ~40–50%.
Matrix snapshot — we also have AA's own HLE column.
| # | Model | Score |
|---|---|---|
| 1 | 93.2 | |
| 2 | 87.2 | |
| 3 | 83.0 | |
| 4 | 69.3 | |
| 5 | 67.6 | |
| 6 | 66.7 | |
| 7 | 66.4 | |
| 8 | 66.0 | |
| 9 | 65.1 | |
| 10 | 65.0 | |
| 11 | 64.7 | |
| 12 | 64.2 | |
| 13 | 64.0 | |
| 14 | 63.9 | |
| 15 | 63.7 | |
| 16 | 63.6 | |
| 17 | 63.1 | |
| 18 | 62.3 | |
| 19 | 61.9 | |
| 20 | 61.9 | |
| 21 | 61.9 | |
| 22 | 61.7 | |
| 23 | 61.5 | |
| 24 | 59.5 | |
| 25 | 59.5 | |
| 26 | 59.1 | |
| 27 | 58.5 | |
| 28 | 57.9 | |
| 29 | 57.8 | |
| 30 | 57.1 | |
| 31 | 56.8 | |
| 32 | 56.6 | |
| 33 | 55.6 | |
| 34 | 54.4 | |
| 35 | 53.7 | |
| 36 | 52.7 | |
| 37 | 52.6 | |
| 38 | 51.4 | |
| 39 | 50.9 | |
| 40 | 50.2 | |
| 41 | 50.2 | |
| 42 | 49.8 | |
| 43 | 49.6 | |
| 44 | 49.3 | |
| 45 | 49.2 | |
| 46 | 49.0 | |
| 47 | 48.9 | |
| 48 | 48.8 | |
| 49 | 48.7 | |
| 50 | 48.4 |