HLE (Humanity's Last Exam)

이 벤치마크는?

인류 최고 수준의 전문 지식을 측정하도록 설계된 최고 난이도 벤치마크. 수학, 과학, 공학 등 50개 이상 분야의 극도로 어려운 문제들로 구성됩니다. 점수는 정답률(%)입니다.

출처: Artificial Analysis
순위모델
#1

Google

Gemini 3.1 Pro

44.7%
#2

OpenAI

GPT-5.4

41.6%
#3

Meta AI

Muse Spark

39.9%
#4

Anthropic

Claude Opus 4.6

36.7%
#5

Google

Gemini 3 Flash

34.7%
#6

Grok

Grok 4.20 (Reasoning)

32.2%
#7

Moonshot AI

Kimi K2.5

29.4%
#8

Anthropic

Claude Opus 4.5

28.4%
#9

Xiaomi

MiMo-V2-Pro

28.3%
#10

MiniMax

MiniMax M2.7

28.1%
#11

Z.ai

GLM-5

27.2%
#12

Qwen

Qwen3.6 Plus

25.7%
#13

Grok

Grok 4.20

24.2%
#14

Google

Gemma 4 31B

22.7%
#15

DeepSeek

DeepSeek V3.2

22.2%
#16

Google

Gemini 2.5 Pro

21.1%
#17

NVIDIA

Nemotron 3 Super

19.2%
#18

MiniMax

MiniMax M2.5

19.1%
#19

Qwen

Qwen3.5 397B A17B

18.8%
#20

Grok

Grok 4.1 Fast (Reasoning)

17.6%
#21

Anthropic

Claude Sonnet 4.5

17.3%
#22

OpenAI

GPT-5.4 Mini

17.1%
#23

Google

Gemini 3.1 Flash Lite

16.2%
#24

OpenAI

GPT-5.4 Nano

14.7%
#25

Baidu

ERNIE 5.0 Thinking

12.7%
#26

Anthropic

Claude Opus 4.1

11.9%
#27

Anthropic

Claude Opus 4

11.7%
#28

Google

Gemini 2.5 Flash

11.1%
#29

Anthropic

Claude Sonnet 4.6

10.8%
#30

Anthropic

Claude Haiku 4.5

9.7%
#31

Anthropic

Claude Sonnet 4

9.6%
#32

OpenAI

GPT-5 Nano

7.6%
#33

Meituan

Longcat Flash Chat

6.0%
#34

OpenAI

GPT-5

5.8%
#35

OpenAI

GPT OSS 120B

5.2%
#36

OpenAI

GPT-5 Mini

5.0%
#37

Grok

Grok 4.1 Fast

5.0%
#38

Meta AI

Llama 4 Maverick

4.8%
#39

Google

Gemini 2.5 Flash Lite

4.6%
#40

OpenAI

GPT-4.1

4.6%
#41

Mistral AI

Mistral Small 4

4.4%
#42

Meta AI

Llama 4 Scout

4.3%
#43

Baidu

ERNIE 4.5 300B A47B

3.5%
#44

Amazon

Nova 2 Lite

3.0%