BenchLeader
Menu
Best LLMsChoose a modelLeaderboardsModelsBenchmarksProvidersReleases12

Humanity's Last Exam

2,500 expert-written questions across 100+ subjects (Scale AI / CAIS).

As of 24 Oct 2026, GPT-6 Astra leads Humanity's Last Exam on BenchLeader with 54.8%, ahead of Claude Fable 5.1 at 46.5%, across 48 model configurations with a published result.

Published by
Scale AI / CAISdata via Epoch AI Benchmarking Hub
Category
Reasoning
Index weight
1.0
Models
48
Data as of
24 Oct 2026
Newest result
7 Oct 2026

CC BY 4.0 — Epoch AI, ‘AI Benchmarking Hub’, epoch.ai/benchmarks. Mirrored boards credit their original publishers.

What the test looks like

Humanity's Last Exam is about 2,500 questions written by subject experts across more than a hundred fields, from classics to chemistry, chosen because frontier models got them wrong at the time of writing. Most are exact-answer or multiple-choice; some include images.

How it is scored

Accuracy on the exact-match or multiple-choice answers, published by Scale AI and the Center for AI Safety. Models are also asked for a confidence, and a separate calibration error is reported.

What to keep in mind

Questions were selected adversarially against 2024-era models, so scores rose quickly as newer models arrived. Some questions have since been disputed or corrected. Scores below 30% are common and still meaningful.

Top 15
  1. 1GPT-6 Astra54.8%
  2. 2Claude Fable 5.1 (xhigh)46.5%
  3. 3Gemini 3.1 Pro46.4%
  4. 4Claude Sonnet 5.5 (xhigh)46.1%
  5. 5Gemini 3.8 Flash44.5%
  6. 6GPT-5.4 Pro44.3%
  7. 7Muse Spark40.6%
  8. 8Gemini 3 Pro37.5%
  9. 9GPT-5.4 (xhigh)36.2%
  10. 10Claude Opus 4.736.2%
  11. 11Claude Opus 4.6 (max)34.4%
  12. 12GPT-5 Pro31.6%
  13. 13Claude Haiku 5.530.1%
  14. 14GPT-5.227.8%
  15. 15GPT-5 (high)25.3%
48 of 48
#
1GPT-6 Astra54.8%68.02026-09-03
2Claude Fable 5.1xhigh46.5%69.52026-09-01
3Gemini 3.1 Pro46.4%60.62026-02-19
4Claude Sonnet 5.5xhigh46.1%67.92026-09-28
5Gemini 3.8 Flash44.5%55.72026-09-02
6GPT-5.4 Pro44.3%63.32026-03-05
7Muse Spark40.6%64.32026-04-08
8Gemini 3 Pro37.5%60.42025-11-18
9GPT-5.4xhigh36.2%64.72026-03-05
10Claude Opus 4.736.2%63.32026-04-16
11Claude Opus 4.6max34.4%57.42026-02-05
12GPT-5 Pro31.6%59.82025-10-07
13Claude Haiku 5.530.1%–2026-10-07
14GPT-5.227.8%58.02025-12-11
15GPT-5high25.3%57.92025-08-07
16GPT-525.3%56.22025-08-07
17Claude Opus 4.525.2%57.92025-11-24
18Kimi K2.524.4%54.52026-01-27
19GPT-5.123.7%56.82025-11-13
20Gemini 2.5 Pro21.6%53.92025-06-05
21o3high20.3%53.92025-04-16
22GPT-5 mini19.4%56.52025-08-07
23o3medium19.2%53.32025-04-16
24Claude Opus 4.619.0%62.62026-02-05
25o4-minihigh18.1%50.12025-04-16
26o4-minimedium14.3%51.92025-04-16
27Claude Sonnet 4.513.7%53.72025-09-29
28Gemini 2.5 Flash12.1%52.72025-04-17
29Claude Opus 4.111.5%51.52025-08-05
30Claude Opus 410.7%52.62025-05-22
31Gemini 3.1 Flash Lite8.6%53.72026-03-03
32GLM-4.58.3%51.72025-08-03
33GLM-4.5-Air8.1%46.62025-07-20
34o1-pro8.1%–2025-03-19
35Claude 3.7 Sonnet8.0%50.02025-02-24
36o18.0%53.12024-12-17
37Claude Sonnet 47.8%50.52025-05-22
38GPT-5.1no reasoning6.8%45.92025-11-13
39Gemini 2.0 Flash Thinkingthinking6.6%44.92025-01-21
40Llama 4 Maverick5.7%42.82025-04-05
41GPT-4.55.4%49.72025-02-27
42GPT-4.15.4%48.32025-04-14
43Gemini 1.5 Pro 0024.6%41.62024-09-24
44Mistral Medium 34.5%44.12025-05-07
45Nova Pro4.4%39.82024-12-03
46Claude 3.5 Sonnet4.1%46.32024-10-22
47Nova Lite3.6%38.32024-12-03
48GPT-4o2.7%43.92024-11-20

More reasoning benchmarks · All LLM benchmarks

Cite as: BenchLeader, “Humanity's Last Exam leaderboard”, data as of 24 Oct 2026, https://www.benchleader.com/benchmarks/hle.

Humanity's Last Exam: questions

What does Humanity's Last Exam measure?
Humanity's Last Exam is about 2,500 questions written by subject experts across more than a hundred fields, from classics to chemistry, chosen because frontier models got them wrong at the time of writing. Most are exact-answer or multiple-choice; some include images. Scores are reported in percent of tasks solved; higher is better.
Which AI model leads Humanity's Last Exam?
GPT-6 Astra leads Humanity's Last Exam with 54.8% as of 24 Oct 2026, ahead of Claude Fable 5.1 at 46.5%.
How many models have Humanity's Last Exam results?
48 model configurations have a Humanity's Last Exam result on BenchLeader, all taken from Scale AI / CAIS via Epoch AI Benchmarking Hub.
Who runs Humanity's Last Exam and how often is it updated?
Humanity's Last Exam is published by Scale AI / CAIS. BenchLeader re-reads the published results once a day and records the date each result was published.
Does Humanity's Last Exam count toward the BenchLeader Index?
Yes. Humanity's Last Exam contributes to the reasoning category of the BenchLeader Index, normalised so that 50 is the average of the evaluated models.