Humanity's Last Exam (HLE)

An extremely challenging benchmark crowd-sourced from thousands of domain experts worldwide, designed to probe the absolute frontier of AI capabilities with questions that even specialists find difficult.

About HLE

Year

2025

Tasks

Expert-level questions

Format

Open-ended and multiple choice

Difficulty

Frontier expert level

HLE represents the hardest public benchmark available, with top models scoring only 10-45%. Questions span advanced mathematics, theoretical physics, philosophy, and other fields at the cutting edge of human knowledge.

Humanity's Last Exam

Leaderboard (88 models)

#1GPT-5.4OpenAI

46

#2GPT-5.3 CodexOpenAI

44

#3GPT-5.2OpenAI

42

#4Gemini 3.1 ProGoogle

40

40

#6Claude Opus 4.6Anthropic

38

#7Gemini 3 Pro Deep ThinkGoogle

32

#8o1-previewOpenAI

32

#9GLM-5 (Reasoning)Zhipu AI

29

#10Qwen3.5 397B (Reasoning)Alibaba

29

#11GPT-5.1-Codex-MaxOpenAI

27

#12GPT-5.1OpenAI

27

#13GPT-5 (high)OpenAI

27

#14GPT-5 (medium)OpenAI

27

#15Kimi K2.5 (Reasoning)Moonshot AI

27

#16GPT-5.2-CodexOpenAI

26

#17o3-proOpenAI

26

#18o3OpenAI

24

#19DeepSeek V3.2 (Thinking)DeepSeek

22

#20Claude Sonnet 4.6Anthropic

21

#21Claude Sonnet 4.5Anthropic

21

#22Claude Opus 4.5Anthropic

20

#23Gemini 3 ProGoogle

20

#24Grok 4.1 FastxAI

20

#25DeepSeekMath V2DeepSeek

18

#26GPT-5 miniOpenAI

16

#27Grok 4xAI

16

#28GLM-4.7Zhipu AI

16

#29Nemotron 3 Ultra 500BNVIDIA

15

#30GLM-4.7-FlashZhipu AI

15

#31DeepSeek Coder 2.0DeepSeek

14

#32MiMo-V2-FlashXiaomi

14

#33DeepSeek-R1DeepSeek

14

#34GLM-5Zhipu AI

13

#35o4-mini (high)OpenAI

13

#36Nemotron 3 Super 100BNVIDIA

13

#37DeepSeek LLM 2.0DeepSeek

12

#38Mistral Large 3Mistral

12

#39Claude 4 SonnetAnthropic

12

#40Mistral Large 2Mistral

12

#41Qwen2.5-72BAlibaba

11

#42DeepSeek V3.2DeepSeek

11

#43Claude 4.1 OpusAnthropic

11

#44Kimi K2.5Moonshot AI

11

#45Claude Haiku 4.5Anthropic

11

#46Nemotron Ultra 253BNVIDIA

11

#47Qwen2.5-1MAlibaba

10

#48Qwen3.5 397BAlibaba

10

#49MiniMax M2.5MiniMax

10

#50DeepSeek V3.1 (Reasoning)DeepSeek

10

#51Mistral 8x7BMistral

8

#52Claude 4.1 Opus ThinkingAnthropic

8

#53Llama 3.1 405BMeta

7

#54Grok Code Fast 1xAI

7

#55Gemini 3 FlashGoogle

6

#56Z-1Z

6

#57Qwen3 235B 2507 (Reasoning)Alibaba

6

#58Claude 3.5 SonnetAnthropic

5

#59Nemotron-4 15BNVIDIA

5

#60Moonshot v1Moonshot AI

5

#61GPT-OSS 120BOpenAI

5

#62Mistral 7B v0.3Mistral

5

#63Llama 4 MaverickMeta

4

#64Nova ProNova AI

4

#65GLM-4.5-AirTsinghua

4

#66Gemini 2.5 ProGoogle

3

#67Llama 4 BehemothMeta

3

#68Gemma 3 27BGoogle

3

#69Grok 3 [Beta]xAI

3

#70GLM-4.5Tsinghua

3

#71Kimi K2Moonshot AI

3

#72Mistral 8x7B v0.2Mistral

3

#73Llama 3 70BMeta

2

#74Claude 3 HaikuAnthropic

2

#75Llama 4 ScoutMeta

2

#76Qwen2.5-VL-32BAlibaba

2

#77MiniMax M1 80kMiniMax

2

#78DeepSeek V3.1DeepSeek

2

#79GPT-4oOpenAI

1

#80Gemini 1.5 ProGoogle

1

#81Gemini 3.1 Flash-LiteGoogle

1

#82Claude 3 OpusAnthropic

1

#83GPT-4 TurboOpenAI

1

#84Gemini 1.0 ProGoogle

1

#85Nemotron 3 Nano 30BNVIDIA

1

#86Gemini 2.5 FlashGoogle

1

#87Qwen3 235B 2507Alibaba

1

#88GPT-OSS 20BOpenAI

1

FAQ

What does HLE measure?

An extremely challenging benchmark crowd-sourced from thousands of domain experts worldwide, designed to probe the absolute frontier of AI capabilities with questions that even specialists find difficult.

Which model scores highest on HLE?

GPT-5.4 by OpenAI currently leads with a score of 46 on HLE.

How many models are evaluated on HLE?

88 AI models have been evaluated on HLE on BenchLM.