τ²-bench

Mean pass^1 across the airline, retail and telecom domains.

Published by

τ²-bench

Category

Agents & tools

Index weight

1.0

Models

8

Data as of

9 Sept 2026

MIT — sierra-research/tau2-bench.

1. 1Qwen3.5 397B-A17B87.9%
2. 2Claude Opus 4.585.3%
3. 3GPT-5.284.8%
4. 4Gemini 3 Flash83.5%
5. 5Gemini 3 Pro82.5%
6. 6GLM-581.0%
7. 7Claude Sonnet 4.576.4%
8. 8GPT-5.261.6%

SearchAll providersAlibabaAnthropicGoogleOpenAIZhipu AIOpen weights8 of 8

#	Model	Score ↓	Run ↕	Index ↕	Published ↕
1	Qwen3.5 397B-A17BAlibabaopen	87.9%	–	–	–
2	Claude Opus 4.5Anthropic	85.3%	–	56.6	–
3	GPT-5.2OpenAI	84.8%	–	55.5	–
4	Gemini 3 FlashGoogle	83.5%	–	58.5	–
5	Gemini 3 ProGoogle	82.5%	–	57.0	–
6	GLM-5Zhipu AIopen	81.0%	–	57.0	–
7	Claude Sonnet 4.5Anthropic	76.4%	–	52.3	–
8	GPT-5.2OpenAI	61.6%	–	50.1	–