benchmarks.site

Public leaderboard · General knowledge & reasoning

Humanity's Last Exam (HLE)

About 2,500 extremely hard questions written by experts in over 100 subjects, from classics to quantum physics, meant to sit at the edge of human knowledge. It is one of the hardest general-knowledge exams for AI.

Dataset

Top 3 right now

  1. 1stGPT 6 Astra54.8
  2. 2ndFable 5.1 (xhigh)46.5
  3. 3rdgemini-3.1-pro-preview (thinking high)46.4

What it measures

About 2,500 expert-written questions across more than 100 subjects, some with images, scored by accuracy on exact-match and multiple-choice answers.

How to read the numbers

Share of questions answered correctly (%) — higher is better.

Results

25 rows · higher is better (Accuracy (%))CSV
1GPT 6 AstraOpenAI54.839.02026-09-03
2Fable 5.1 (xhigh)Anthropic46.520.02026-09-01
3gemini-3.1-pro-preview (thinking high)Google DeepMind46.451.02026-02-19
4Sonnet 5.5 (xhigh)Anthropic46.12026-09-28
5Gemini 3.8 FlashGoogle DeepMind44.551.02026-09-02
6gpt-5.4-pro-2026-03-05OpenAI44.338.02026-03-05
7Muse SparkMeta AI40.650.02026-04-08
8gemini-3-pro-previewGoogle DeepMind37.557.02025-11-18
9gpt-5.4-2026-03-05 (xhigh thinking)OpenAI36.242.02026-03-05
10claude-opus-4-7Anthropic36.247.02026-04-16
11claude-opus-4-6-thinking-maxAnthropic34.446.02026-02-05
12gpt-5-pro-2025-10-06OpenAI31.649.02025-10-07
13Claude Haiku 5.5Anthropic30.12026-10-07
14gpt-5.2-2025-12-11OpenAI27.845.02025-12-11
15gpt-5-2025-08-07OpenAI25.350.02025-08-07
16gpt-5-2025-08-07OpenAI25.350.02025-08-07
17claude-opus-4-5-20251101-thinkingAnthropic25.255.02025-11-24
18kimi-k2.5Moonshot24.467.02026-01-27
19gpt-5.1-thinkingOpenAI23.755.02025-11-13
20gemini-2.5-pro-preview-06-05Google DeepMind21.672.02025-06-05
21o3 (high) (April 2025)OpenAI20.334.02025-04-16
22gpt-5-mini-2025-08-07OpenAI19.465.02025-08-07
23o3 (medium) (April 2025)OpenAI19.239.02025-04-16
24claude-opus-4-6 (Non-Thinking)Anthropic19.044.02026-02-05
25Gemini 2.5 Pro Experimental (March 2025)Google DeepMind18.271.02025-03-25

Note: Externally reported results mirrored by Epoch AI (hle_external.csv); accuracy converted to % (1 decimal), calibration error copied as stored. Audits have found wrong reference answers in part of the set.

Source: Epoch AI (Benchmarking Hub) · epoch.ai · results as of 2026-11-17 · retrieved 2026-11-17 · license: CC-BY-4.0

Markdown · More General knowledge & reasoning