τ²-bench
Mean pass^1 across the airline, retail and telecom domains.
As of 25 Oct 2026, Qwen3.5 397B A17B leads τ²-bench on BenchLeader with 87.9%, ahead of Claude Opus 4.5 at 85.3%, across 8 model configurations with a published result.
- Published by
- τ²-bench
- Category
- Agents & tools
- Index weight
- 1.0
- Models
- 8
- Data as of
- 25 Oct 2026
- Newest result
- 2 Mar 2026
The publisher has posted no new results on this board since 2 Mar 2026. BenchLeader re-reads it every day, so newer models appear here as soon as they are published.
MIT — sierra-research/tau2-bench.
What the test looks like
τ²-bench simulates customer-service conversations in airline, retail and telecom domains, where the model must follow policy, use tools and coordinate with a simulated user who also takes actions.
How it is scored
Mean pass^1 across domains: the share of conversations ending in the correct final state on the first try.
What to keep in mind
The simulated user is itself a model, so its behaviour affects results. Policy documents are long, and scores depend on how well the model follows them, not on general knowledge.
- 1Qwen3.5 397B A17B (thinking)87.9%
- 2Claude Opus 4.5 (high)85.3%
- 3GPT-5.2 (high)84.8%
- 4Gemini 3 Flash (high)83.5%
- 5Gemini 3 Pro (high)82.5%
- 6GLM-5 (thinking)81.0%
- 7Claude Sonnet 4.576.4%
- 8GPT-5.2 (none)61.6%
| # | ||||
|---|---|---|---|---|
| 1 |  Qwen3.5 397B A17Bthinkingopen ↗ | 87.9% | 55.1 | 2026-03-02 |
| 2 |  Claude Opus 4.5high | 85.3% | 59.2 | 2026-02-26 |
| 3 |  GPT-5.2high | 84.8% | 56.0 | 2026-02-26 |
| 4 |  Gemini 3 Flashhigh | 83.5% | 58.3 | 2026-03-02 |
| 5 | Gemini 3 Prohigh | 82.5% | 60.0 | 2026-03-02 |
| 6 |  GLM-5thinkingopen ↗ | 81.0% | 58.9 | 2026-03-02 |
| 7 | Claude Sonnet 4.5 | 76.4% | 53.7 | 2026-02-26 |
| 8 | GPT-5.2no reasoning | 61.6% | 48.8 | 2026-02-26 |
More agents & tools benchmarks · All LLM benchmarks
Cite as: BenchLeader, “τ²-bench leaderboard”, data as of 25 Oct 2026, https://www.benchleader.com/benchmarks/tau2_bench.
τ²-bench: questions
- What does τ²-bench measure?
- τ²-bench simulates customer-service conversations in airline, retail and telecom domains, where the model must follow policy, use tools and coordinate with a simulated user who also takes actions. Scores are reported in percent of tasks solved; higher is better.
- Which AI model leads τ²-bench?
- Qwen3.5 397B A17B leads τ²-bench with 87.9% as of 25 Oct 2026, ahead of Claude Opus 4.5 at 85.3%.
- How many models have τ²-bench results?
- 8 model configurations have a τ²-bench result on BenchLeader, all taken from τ²-bench.
- Who runs τ²-bench and how often is it updated?
- τ²-bench is published by τ²-bench. BenchLeader re-reads the published results once a day and records the date each result was published.
- Does τ²-bench count toward the BenchLeader Index?
- Yes. τ²-bench contributes to the agents & tools category of the BenchLeader Index, normalised so that 50 is the average of the evaluated models.