tau2-bench Airline
Airline customer-service agent following policy and using booking tools with a simulated user. Few recent models are run on it.
| Category | Agents |
|---|---|
| Metric | pass^1 (%) |
| Source | tau2-bench (original) |
| License | MIT |
| Benchmark page | taubench.com |
| Status | Scored |
| Weight | 0.7 within Agents |
| Fit | Discrimination 0.68, noise 0.36 |
Results
| # | Model | Provider | Result | Date |
|---|---|---|---|---|
| 1 | Claude Opus 4.5 High | Anthropic | 84.0% | 2026-05-05 |
| 2 | GPT-5.2 High | OpenAI | 83.0% | 2026-05-05 |
| 3 | GLM-5 | Zhipu AI | 82.5% | 2026-02-27 |
| 4 | Gemini 3 Flash Preview High | 82.5% | 2026-03-02 | |
| 5 | Qwen3.5 397B-A17B | Alibaba | 81.5% | 2026-02-27 |
| 6 | Gemini 3 Pro Preview High | 80.5% | 2026-03-02 | |
| 7 | Claude Sonnet 4.5 | Anthropic | 72.0% | 2026-02-24 |
| 8 | GPT-5.2 Non-reasoning | OpenAI | 52.5% | 2026-05-26 |