TAU2

이 벤치마크는?

도구 사용 능력을 평가하는 에이전트 벤치마크. 실제 사용자 작업을 시뮬레이션한 멀티스텝 태스크에서 AI의 도구 활용 및 계획 능력을 측정합니다. 점수는 태스크 성공률(%)입니다.

출처: Artificial Analysis
순위모델
#1

Z.ai

GLM-5

98.2%
#2

Qwen

Qwen3.6 Plus

97.7%
#3

Moonshot AI

Kimi K2.5

95.9%
#4

Google

Gemini 3.1 Pro

95.6%
#5

MiniMax

MiniMax M2.5

95.3%
#6

Xiaomi

MiMo-V2-Pro

95.0%
#7

Grok

Grok 4.1 Fast (Reasoning)

93.3%
#8

Grok

Grok 4.20 (Reasoning)

93.0%
#9

Anthropic

Claude Opus 4.6

92.1%
#10

Meta AI

Muse Spark

91.5%
#11

DeepSeek

DeepSeek V3.2

90.6%
#12

Anthropic

Claude Opus 4.5

89.5%
#13

OpenAI

GPT-5.4

87.1%
#14

MiniMax

MiniMax M2.7

84.8%
#15

Baidu

ERNIE 5.0 Thinking

83.9%
#16

Qwen

Qwen3.5 397B A17B

83.9%
#17

Google

Gemini 3 Flash

80.4%
#18

Meituan

Longcat Flash Chat

79.5%
#19

Anthropic

Claude Sonnet 4.6

78.9%
#20

Anthropic

Claude Sonnet 4.5

78.1%
#21

Anthropic

Claude Opus 4

73.4%
#22

Anthropic

Claude Opus 4.1

71.4%
#23

NVIDIA

Nemotron 3 Super

67.8%
#24

Anthropic

Claude Sonnet 4

64.6%
#25

Grok

Grok 4.1 Fast

63.7%
#26

Amazon

Nova 2 Lite

62.0%
#27

Google

Gemma 4 31B

59.9%
#28

Grok

Grok 4.20

59.9%
#29

Anthropic

Claude Haiku 4.5

54.7%
#30

Google

Gemini 2.5 Pro

54.1%
#31

OpenAI

GPT-5.4 Nano

52.6%
#32

OpenAI

GPT-4.1

47.1%
#33

OpenAI

GPT OSS 120B

45.0%
#34

OpenAI

GPT-5.4 Mini

36.5%
#35

OpenAI

GPT-5 Mini

31.9%
#36

Google

Gemini 2.5 Flash

31.6%
#37

Google

Gemini 3.1 Flash Lite

31.3%
#38

Google

Gemini 2.5 Flash Lite

30.4%
#39

OpenAI

GPT-5 Nano

30.4%
#40

Meta AI

Llama 4 Maverick

17.8%
#41

Meta AI

Llama 4 Scout

15.5%
#42

Baidu

ERNIE 4.5 300B A47B

0.0%
#43

OpenAI

GPT-5

0.0%