τ²-bench (domain average)
Agents · 2025-06-09
Release-reported single τ²-bench score that averages task success across the airline, retail and telecom customer-service domains instead of reporting each domain. Airline-domain task fixes and user-simulator settings vary by publisher and are recorded in score notes; domain-specific results use the per-domain τ²-bench rows.
Top models (higher is better)
| Model | Score |
|---|---|
| Opus 4.5 | 91.6 |
| GPT-5.2 | 87.1 |
| Qwen3.5 397B A17B | 86.7 |
| Gemini 3 Pro Preview | 85.4 |
| Qwen3 Max Thinking (2026-01-23) | 84.6 |
| Qwen3.5 Plus (2026-02-15) | 82.7 |
| Qwen3.5 35B A3B | 81.2 |
| Qwen3.5-Omni-Plus | 81.0 |
| Qwen3.5-4B | 79.9 |
| Qwen3.5 122B A10B | 79.5 |
| Qwen3.5-9B | 79.1 |
| Qwen3.5-27B | 79.0 |
| Qwen3.5-Omni-Flash | 78.0 |
| Kimi K2.5 | 77.0 |
| GPT-5 Mini | 69.8 |
| Qwen3-235B-A22B | 58.5 |
| Qwen3-Next-80B-A3B-Thinking | 57.4 |
| Qwen3.5-2B | 48.8 |
| Qwen3 4B Thinking 2507 | 43.2 |
| Qwen3 30B A3B Thinking 2507 | 41.9 |
| Qwen3.5-0.8B | 11.6 |