tau2-bench Banking Knowledge

Customer-service agent for a bank that must search a knowledge base and use tools while talking to a simulated user. Standard Sierra scaffold, GPT-5.2 user simulator.

CategoryAgents
Metricpass^1 (%)
Sourcetau2-bench (original)
LicenseMIT
Benchmark pagetaubench.com
StatusScored
Weight1.0 within Agents
FitDiscrimination 0.75, noise 0.29

Results

#ModelProviderResultDate
1 Qwen3.8 Max Extra High Alibaba 55.2% 2026-08-03
2 Claude Opus 5 Max Anthropic 48.7% 2026-08-03
3 Grok 4.5 High xAI 47.9% 2026-07-24
4 GPT-5.6 Sol Extra High OpenAI 46.9% 2026-07-22
5 GPT-5.5 Extra High OpenAI 44.6% 2026-07-22
6 Muse Spark 1.1 Extra High Meta 40.5% 2026-07-23
7 Claude Opus 4.7 Max Anthropic 40.2% 2026-07-23
8 Claude Fable 5 Max Anthropic 39.7% 2026-07-23
9 Claude Opus 4.8 Max Anthropic 39.7% 2026-07-23
10 GPT-5.4 Extra High OpenAI 39.4% 2026-05-06
11 Kimi K3 Max Moonshot AI 37.1% 2026-07-24
12 GLM-5.2 Extra High Zhipu AI 37.1% 2026-07-24
13 GPT-5.2 High OpenAI 32.2% 2026-05-05
14 Claude Opus 4.6 Max Anthropic 27.3% 2026-05-06
15 Gemini 3 Flash Preview High Google 27.3% 2026-03-02
16 Gemini 3.1 Pro Preview High Google 26.0% 2026-05-05
17 Claude Sonnet 4.5 Anthropic 25.3% 2026-02-24
18 Inkling Max Thinking Machines 25.0% 2026-07-24
19 Claude Opus 4.5 High Anthropic 24.7% 2026-05-05
20 Gemini 3 Pro Preview High Google 18.0% 2026-03-02
21 Grok 4.20 (Reasoning) High xAI 18.0% 2026-05-05
22 Grok 4 Fast High xAI 15.7% 2026-05-05
23 Gemini 2.5 Pro High Google 13.7% 2026-05-05
24 Grok 4.1 Fast High xAI 13.1% 2026-05-05
25 GPT-5.2 Non-reasoning OpenAI 12.6% 2026-05-26
26 GLM-5 Zhipu AI 9.8% 2026-02-27
27 Qwen3.5 397B-A17B Alibaba 9.8% 2026-02-27