Humanity's Last Exam
Expert-written questions spanning many specialist fields. Aggregated from Epoch AI's AI Benchmarking Hub (hle_external.csv, column 'Accuracy'), used under CC-BY 4.0. The Known Good aggregates published results; it runs no evaluations.
48 models scored
Higher is better
Published by Epoch AI
CC-BY 4.0
Humanity's Last Exam
Humanity's Last Exam as measured and published by Epoch AI (CC-BY 4.0). Ingested, normalised only where it feeds an index, and never re-run by us. Higher is better.
💡 A lightbulb marks a model its provider publishes as a reasoning model
The Known Good
Every model we hold a Humanity's Last Exam score for
All 48 of 48 models
| # | Model | Creator | Humanity's Last Exam | Known Good Index | Measured |
|---|---|---|---|---|---|
| 1 |
OpenAI: GPT-6 Astra
💡 |
OpenAI | 54.8 | — | 8 Oct 2026 |
| 2 |
Anthropic: Claude Fable 5.1
💡 |
Anthropic | 46.5 | — | 8 Oct 2026 |
| 3 |
Google: Gemini 3.1 Pro Preview
💡 |
46.4 | 93 | 8 Oct 2026 | |
| 4 |
Google: Gemini 3.8 Flash
💡 |
44.5 | — | 8 Oct 2026 | |
| 5 |
OpenAI: GPT-5.4 Pro
💡 |
OpenAI | 44.3 | — | 8 Oct 2026 |
| 6 |
muse-spark
|
Meta | 40.6 | — | 8 Oct 2026 |
| 7 |
gemini-3-pro-preview
|
37.5 | 81 | 8 Oct 2026 | |
| 8 |
OpenAI: GPT-5.4
💡 |
OpenAI | 36.2 | 89 | 8 Oct 2026 |
| 9 |
Anthropic: Claude Opus 4.7
💡 |
Anthropic | 36.2 | 90 | 8 Oct 2026 |
| 10 |
Anthropic: Claude Opus 4.6
💡 |
Anthropic | 34.4 | 87 | 8 Oct 2026 |
| 11 |
OpenAI: GPT-5 Pro
💡 |
OpenAI | 31.6 | — | 8 Oct 2026 |
| 12 |
OpenAI: GPT-5.2
💡 |
OpenAI | 27.8 | 79 | 8 Oct 2026 |
| 13 |
OpenAI: GPT-5
💡 |
OpenAI | 25.3 | 77 | 8 Oct 2026 |
| 14 |
claude-opus-4-5-20251101-thinking
|
Anthropic | 25.2 | — | 8 Oct 2026 |
| 15 |
MoonshotAI: Kimi K2.5
💡 |
Moonshot AI | 24.4 | — | 8 Oct 2026 |
| 16 |
OpenAI: GPT-5.1
💡 |
OpenAI | 23.7 | 73 | 8 Oct 2026 |
| 17 |
Google: Gemini 2.5 Pro Preview 06-05
💡 |
21.6 | — | 8 Oct 2026 | |
| 18 |
OpenAI: o3
💡 |
OpenAI | 20.3 | 72 | 8 Oct 2026 |
| 19 |
OpenAI: GPT-5 Mini
💡 |
OpenAI | 19.4 | 66 | 8 Oct 2026 |
| 20 |
Gemini 2.5 Pro Experimental (March 2025)
|
18.2 | — | 8 Oct 2026 | |
| 21 |
OpenAI: o4 Mini
💡 |
OpenAI | 18.1 | — | 8 Oct 2026 |
| 22 |
Google: Gemini 2.5 Pro Preview 05-06
💡 |
17.8 | — | 8 Oct 2026 | |
| 23 |
Anthropic: Claude Opus 4.5
💡 |
Anthropic | 14.2 | 71 | 8 Oct 2026 |
| 24 |
claude-sonnet-4-5-20250929-thinking
|
Anthropic | 13.7 | — | 8 Oct 2026 |
| 25 |
Google: Gemini 2.5 Flash
💡 |
12.1 | — | 8 Oct 2026 | |
| 26 |
claude-opus-4-1-20250805-thinking
|
Anthropic | 11.5 | — | 8 Oct 2026 |
| 27 |
Gemini 2.5 Flash Preview (May 2025)
|
11.0 | — | 8 Oct 2026 | |
| 28 |
Anthropic: Claude Opus 4
💡 |
Anthropic | 10.7 | — | 8 Oct 2026 |
| 29 |
Google: Gemini 3.1 Flash Lite Preview
💡 |
8.6 | — | 8 Oct 2026 | |
| 30 |
Z.ai: GLM 4.5
💡 |
Z.ai | 8.3 | — | 8 Oct 2026 |
| 31 |
Z.ai: GLM 4.5 Air
💡 |
Z.ai | 8.1 | — | 8 Oct 2026 |
| 32 |
OpenAI: o1-pro
💡 |
OpenAI | 8.1 | — | 8 Oct 2026 |
| 33 |
Claude 3.7 Sonnet (Thinking)
|
Anthropic | 8.0 | — | 8 Oct 2026 |
| 34 |
OpenAI: o1
💡 |
OpenAI | 8.0 | 69 | 8 Oct 2026 |
| 35 |
Anthropic: Claude Opus 4.1
💡 |
Anthropic | 7.9 | 49 | 8 Oct 2026 |
| 36 |
Anthropic: Claude Sonnet 4
💡 |
Anthropic | 7.8 | — | 8 Oct 2026 |
| 37 |
claude-sonnet-4-5-20250929
|
Anthropic | 7.5 | 66 | 8 Oct 2026 |
| 38 |
gpt-5.1-instant
|
OpenAI | 6.8 | — | 8 Oct 2026 |
| 39 | GD Gemini 2.0 Flash Thinking (January 2025) | Google DeepMind,Google | 6.6 | — | 8 Oct 2026 |
| 40 |
Meta: Llama 4 Maverick
|
Meta | 5.7 | — | 8 Oct 2026 |
| 41 |
gpt-4.5-preview
|
OpenAI | 5.4 | 54 | 8 Oct 2026 |
| 42 |
OpenAI: GPT-4.1
|
OpenAI | 5.4 | 46 | 8 Oct 2026 |
| 43 |
gemini-1.5-pro-002
|
4.6 | — | 8 Oct 2026 | |
| 44 |
Mistral: Mistral Medium 3
|
Mistral | 4.5 | — | 8 Oct 2026 |
| 45 | Az Nova Pro | Amazon | 4.4 | — | 8 Oct 2026 |
| 46 |
Claude 3.5 Sonnet (October 2024)
|
Anthropic | 4.1 | — | 8 Oct 2026 |
| 47 | Az Nova Lite | Amazon | 3.6 | — | 8 Oct 2026 |
| 48 |
OpenAI: GPT-4o
|
OpenAI | 2.7 | 27 | 8 Oct 2026 |
The Known Good
Provenance. These figures are published by Epoch AI and redistributed here under CC-BY 4.0.
The publisher's own page is here. We did not run this evaluation and we do not adjust the published numbers — where a score feeds an index it is
min-max normalised against every other model holding the same evaluation, and nothing else is done to it.
This evaluation is one of the seven components of the
Known Good Index. Full licence detail is on attribution, and every score here is
in the CSV download.