Humanity's Last Exam
Integrity rank #3 of 61 · 37 models scored · top score 43.74 · Gemini 3.1 Pro
Frontier expert knowledge/reasoning across 100+ subjects
Strongest on discrimination, weakest on saturation headroom. Caveat: the public test set does not rule out training-data contamination.
Reported scores — protocols may differ. How we compare
| # | Model | Score | Evidence | Measured |
|---|---|---|---|---|
| 1 | Gemini 3.1 Pro | 43.74 | unverifiedT2 | 2026-02-19 |
| 2 | GPT-5.4 Pro | 41.51 | unverifiedT2 | 2026-03-05 |
| 3 | Muse Spark | 37.56 | unverifiedT2 | 2026-04-08 |
| 4 | Gemini 3 Pro | 34.37 | unverifiedT2 | 2025-11-18 |
| 5 | GPT-5.4 | 33.03 | unverifiedT2 | 2026-03-05 |
| 6 | Claude Opus 4.7 | 32.98 | unverifiedT2 | 2026-04-16 |
| 7 | Claude Opus 4.6 | 31.13 | unverifiedT2 | 2026-02-05 |
| 8 | GPT-5 Pro | 28.19 | unverifiedT2 | 2025-10-07 |
| 9 | GPT-5.2 | 24.16 | unverifiedT2 | 2025-12-11 |
| 10 | GPT-5 | 21.55 | unverifiedT2 | 2025-08-07 |
| 11 | Claude Opus 4.5 | 21.43 | unverifiedT2 | 2025-11-24 |
| 12 | Kimi K2.5 | 20.56 | unverifiedT2 | 2026-02-02 |
| 13 | GPT-5.1 | 19.83 | unverifiedT2 | 2025-11-13 |
| 14 | Gemini 2.5 Pro (Jun 2025) | 17.69 | unverifiedT2 | 2025-06-05 |
| 15 | o3 | 16.3 | unverifiedT2 | 2024-12-20 |
| 16 | GPT-5 mini | 15.38 | unverifiedT2 | 2025-08-07 |
| 17 | Gemini 2.5 Pro (Mar 2025) | 14.03 | unverifiedT2 | 2025-03-25 |
| 18 | o4-mini | 13.95 | unverifiedT2 | 2025-04-16 |
| 19 | Gemini 2.5 Pro (May 2025) | 13.66 | unverifiedT2 | 2025-05-06 |
| 20 | Claude Sonnet 4.5 | 9.37 | unverifiedT2 | 2025-09-29 |
| 21 | Gemini 2.5 Flash (Apr 2025) | 7.65 | unverifiedT2 | 2025-04-17 |
| 22 | Claude Opus 4.1 | 7.06 | unverifiedT2 | 2025-08-05 |
| 23 | Gemini 2.5 Flash (May 2025) | 6.47 | unverifiedT2 | 2025-05-20 |
| 24 | Claude Opus 4 | 6.22 | unverifiedT2 | 2025-05-22 |
| 25 | Gemini 3.1 Flash-Lite | 4.03 | unverifiedT2 | 2026-03-03 |
| 26 | Claude 3.7 Sonnet | 3.4 | unverifiedT2 | 2025-02-24 |
| 27 | o1 | 3.32 | unverifiedT2 | 2024-12-05 |
| 28 | Claude Sonnet 4 | 3.11 | unverifiedT2 | 2025-05-22 |
| 29 | Gemini 2.0 Flash Thinking (Jan 2025) | 1.85 | unverifiedT2 | 2025-01-21 |
| 30 | Llama 4 Maverick | 0.92 | unverifiedT2 | 2025-04-05 |
| 31 | GPT-4.5 | 0.67 | unverifiedT2 | 2025-02-27 |
| 32 | GPT-4.1 | 0.63 | unverifiedT2 | 2025-04-14 |
| 33 | Amazon Nova Pro | 0 | unverifiedT2 | 2024-12-03 |
| 34 | Claude 3.5 Sonnet (October 2024) | 0 | unverifiedT2 | 2024-10-22 |
| 35 | Mistral Medium 3 | 0 | unverifiedT2 | 2025-05-07 |
| 36 | GPT-4o (Nov 2024) | 0 | unverifiedT2 | 2024-05-13 |
| 37 | Gemini 1.5 Pro (Sept 2024) | 0 | unverifiedT2 | 2024-09-24 |
Continue from this leaderboard into the tightest head-to-head reads.
Weighted composite. Each dimension opens to the source-backed sub-signals it is computed from.
Claims drawn from cited facts, not live model generation.
This benchmark scores 37 models, with a spread of 43.74 points between the best and worst scores, indicating a wide range that supports genuine capability separation between models.
2 cited facts
This benchmark receives an integrity score of 100 (grade A), placing it 3rd among 51 benchmarks. The weakest component in its integrity breakdown is saturation, indicating that the ceiling is crowded.
5 cited facts
The benchmark is not saturated, as indicated by the saturation flag being false. 56.26 points of open scale sit above the 43.74 leader — separation at the top is limited by results, not by the test, so leaderboard movement here is measurement rather than ceiling artifact. Only a single model clusters near the top, meaning that small differences in scores still reflect genuine capability gaps.
4 cited facts
Directly comparable means comparable to each other under this setup; the moment a score from this table is set against an externally reported figure, that guarantee has already been spent. The test set is public, so high task performance warrants heightened scrutiny for possible contamination, especially given the expected public leakage noted in the contamination history. Sensitivity is high, as task performance shows large swings between text-only and browsing-enabled evaluation modes.
4 cited facts
Breadth from many contributors and multi-stage review buys coverage, not coherence; the score averages very different subjects and formats, so treat it as a survey result rather than a measure of any particular skill. Questions chosen because they stump current models make the frontier a moving target. As models improve, the test's meaning shifts with them, so scores from different times are not measuring quite the same thing. The most important caveat is that the benchmark is limited to static, short-answer and multiple-choice formats, so it cannot evaluate agentic tasks or real-world multi-step reasoning, and the public portion is vulnerable to contamination.
4 cited facts
HLE (Humanity's Last Exam): Humanity's Last Exam — Frontier expert knowledge/reasoning across 100+ subjects; scored as Accuracy + calibration/confidence-error.
Gemini 3.1 Pro leads HLE at 43.74. The full leaderboard above lists every recorded measurement, not just the headline number.
37 models have recorded scores on HLE, spanning a score spread of 43.74.
tensor.news grades HLE A for integrity (score 100/100), ranking #3 of 61 benchmarks we assess across discrimination, saturation, contamination resistance, harness comparability, and freshness.
No — HLE still has headroom and continues to discriminate between models rather than bunching them at the ceiling.
Its test set is public, so contamination risk is on the table: a high score may partly reflect training-data overlap rather than capability.
Scores are reported under a consistent harness, so comparisons on HLE are reasonably apples-to-apples.
Every HLE measurement is source-backed and tagged reproduced, self-reported, or unverified; the underlying record cites agi.safe.ai.