τ²-BenchAgents

TAU2-Bench is the second-generation version of TAU-bench, introducing more complex multi-turn conversations and improved evaluation methods. It adds richer user models, more detailed scoring criteria, and more varied task templates for evaluating customer-service agents.

Ranking Sources

#ModelScore

1

Claude Opus 4.5Anthropic

91.6%

2

Gemini 3 ProGoogle

90.7%

3

GLM-5Zhipu AI

89.7%

4

Claude Sonnet 4.5Anthropic

88.1%

5

GLM-4.7Zhipu AI

87.4%

6

GPT-5.2OpenAI

87.1%

7

Qwen3.5-397B-A17BAlibaba

86.7%

8

DeepSeek-V3.2DeepSeek

85.3%

9

GPT-5.1OpenAI

82.7%

10

GPT-5OpenAI

82.4%

21 models total

Back to benchmarks