τ²-bench (Retail)
Agents · 2025-06-09
The retail domain of τ²-bench evaluates an agent handling customer-service tasks against a policy document and a set of tools, with a simulated user. Scores report end-to-end task success as pass^1 over independent trials.
Top models (higher is better)
| Model | Score |
|---|---|
| Opus 4.6 | 91.9 |
| Sonnet 4.6 | 91.7 |
| Gemini 3.1 Pro Preview | 90.8 |
| Qwen3.5 397B A17B | 84.4 |
| GPT-5.2 | 82.0 |
| Opus 4.5 | 79.6 |
| Gemini 3 Flash Preview | 76.8 |
| GLM-5 | 73.7 |
| Sonnet 4.5 | 72.4 |