tau2-bench Retail
Retail customer-service agent handling orders, returns and exchanges with a simulated user. Close to saturation.
| Category | Agents |
|---|---|
| Metric | pass^1 (%) |
| Source | tau2-bench (original) |
| License | MIT |
| Benchmark page | taubench.com |
| Status | Scored |
| Weight | 0.6 within Agents |
| Fit | Discrimination 0.13, noise 0.34 |
Results
| # | Model | Provider | Result | Date |
|---|---|---|---|---|
| 1 | Qwen3.5 397B-A17B | Alibaba | 84.4% | 2026-02-27 |
| 2 | GPT-5.2 High | OpenAI | 81.6% | 2026-05-05 |
| 3 | Claude Opus 4.5 High | Anthropic | 79.6% | 2026-05-05 |
| 4 | Gemini 3 Flash Preview High | 76.8% | 2026-03-02 | |
| 5 | Gemini 3 Pro Preview High | 75.9% | 2026-03-02 | |
| 6 | GPT-5.2 Non-reasoning | OpenAI | 75.0% | 2026-05-26 |
| 7 | GLM-5 | Zhipu AI | 73.7% | 2026-02-27 |
| 8 | Claude Sonnet 4.5 | Anthropic | 72.4% | 2026-02-24 |