Glossary
τ-bench Verified (airline)
τ-bench Verified (airline), on the OpenRouter Benchmarks leaderboard (CC BY 4.0, latest version we checked), simulates an airline support agent that uses systems and must follow company policy until the end of the conversation. The score is the share of tasks completed within the rules.
What it measures
τ-bench Verified (airline), on the OpenRouter Benchmarks leaderboard (CC BY 4.0, latest version we checked), simulates an airline support agent that uses systems and must follow company policy until the end of the conversation. The score is the share of tasks completed within the rules.
When it matters
It matters for agents that run tasks with tools. It shows whether the model follows policy across several steps, like an assistant that looks up an internal rule and finishes the request without inventing an exception.
How to read it
The score is the share of conversations in which the agent finished the task within the rules, on the OpenRouter Benchmarks leaderboard, in the latest version we checked. In real support, breaking a policy halfway through is costly.
Who leads
- Claude Fable 5 Anthropic82%
- DeepSeek V4 Flash 0423 DeepSeek81.7%
- Gemini 3.7 Flash Google80.7%
- Qwen3.5 397B A17B Alibaba80.2%
- GPT-5.1 OpenAI80%
- Ling 3.1 Flash Ant Group79.3%
- Claude Fable 5.1 Anthropic79.2%
- Qwen3.6 27B Alibaba79%
- MiMo V2.6 Pro Xiaomi78.9%
- Claude Sonnet 5.5 Anthropic78.8%
Source: OpenRouter Benchmarks, latest version we checked.