τ³-bench Banking
pass^1 on the banking domain, which Sierra added after the original three. A support agent working a bank's tools against a simulated customer. Sierra's own run; Artificial Analysis' is in the index.
As of 21 Oct 2026, Qwen3.8 Max leads τ³-bench Banking on BenchLeader with 55.1%, ahead of Claude Opus 5 at 48.7%, across 26 model configurations with a published result.
- Published by
- τ²-bench
- Category
- Agents & tools
- Index weight
- Reference only
- Models
- 26
- Data as of
- 21 Oct 2026
MIT — sierra-research/tau2-bench.
- 1Qwen3.8 Max (max)55.1%
- 2Claude Opus 5 (max)48.7%
- 3Grok 4.5 (high)47.9%
- 4GPT-5.6 Sol (xhigh)46.9%
- 5GPT-5.5 (xhigh)44.6%
- 6Muse Spark 1.1 (xhigh)40.5%
- 7Claude Opus 4.7 (max)40.2%
- 8Claude Fable 5 (max)39.7%
- 9Claude Opus 4.8 (max)39.7%
- 10GPT-5.4 (xhigh)39.4%
- 11Kimi K3 (max)37.1%
- 12GLM 5.2 (xhigh)37.1%
- 13GPT-5.2 (high)32.2%
- 14Gemini 3 Flash (high)27.3%
- 15Claude Opus 4.6 (max)27.3%
26 of 26
| # | ||||
|---|---|---|---|---|
| 1 |  Qwen3.8 Maxmax | 55.1% | 63.3 | 2026-08-04 |
| 2 |  Claude Opus 5max | 48.7% | 68.9 | 2026-08-04 |
| 3 |  Grok 4.5high | 47.9% | 59.3 | 2026-08-04 |
| 4 |  GPT-5.6 Solxhigh | 46.9% | 66.6 | 2026-08-04 |
| 5 | GPT-5.5xhigh | 44.6% | 66.1 | 2026-05-05 |
| 6 |  Muse Spark 1.1xhigh | 40.5% | 61.6 | 2026-08-04 |
| 7 | Claude Opus 4.7max | 40.2% | 62.4 | 2026-05-05 |
| 8 | Claude Fable 5max | 39.7% | 69.4 | 2026-08-04 |
| 9 | Claude Opus 4.8max | 39.7% | 63.4 | 2026-08-04 |
| 10 | GPT-5.4xhigh | 39.4% | 64.8 | 2026-03-25 |
| 11 |  Kimi K3maxopen ↗ | 37.1% | 64.8 | 2026-08-04 |
| 12 |  GLM 5.2xhighopen ↗ | 37.1% | – | 2026-08-04 |
| 13 | GPT-5.2high | 32.2% | 56.0 | 2026-02-26 |
| 14 |  Gemini 3 Flashhigh | 27.3% | 58.4 | 2026-03-02 |
| 15 | Claude Opus 4.6max | 27.3% | 57.5 | 2026-05-05 |
| 16 | Gemini 3.1 Prohigh | 26.0% | 60.3 | 2026-05-05 |
| 17 | Claude Sonnet 4.5 | 25.3% | 53.7 | 2026-02-26 |
| 18 | Claude Opus 4.5high | 24.7% | 59.2 | 2026-02-26 |
| 19 | Gemini 3 Prohigh | 18.0% | 60.1 | 2026-03-02 |
| 20 | Grok 4.2high | 18.0% | – | 2026-05-05 |
| 21 | Grok 4 Fasthigh | 15.7% | – | 2026-05-05 |
| 22 | Gemini 2.5 Prohigh | 13.7% | – | 2026-05-05 |
| 23 | Grok 4.1 Fasthigh | 13.1% | – | 2026-05-05 |
| 24 | GPT-5.2no reasoning | 12.6% | 48.9 | 2026-02-26 |
| 25 | GLM-5thinkingopen ↗ | 9.8% | 59.0 | 2026-03-02 |
| 26 | Qwen3.5 397B A17Bthinkingopen ↗ | 9.8% | 55.2 | 2026-03-02 |
More agents & tools benchmarks · All LLM benchmarks
Cite as: BenchLeader, “τ³-bench Banking leaderboard”, https://www.benchleader.com/benchmarks/tau3_banking, data as of 21 Oct 2026.
τ³-bench Banking: questions
- What does τ³-bench Banking measure?
- pass^1 on the banking domain, which Sierra added after the original three. A support agent working a bank's tools against a simulated customer. Sierra's own run; Artificial Analysis' is in the index. Scores are reported in percent of tasks solved; higher is better.
- Which AI model leads τ³-bench Banking?
- Qwen3.8 Max leads τ³-bench Banking with 55.1% as of 21 Oct 2026, ahead of Claude Opus 5 at 48.7%.
- How many models have τ³-bench Banking results?
- 26 model configurations have a τ³-bench Banking result on BenchLeader, all taken from τ²-bench.
- Who runs τ³-bench Banking and how often is it updated?
- τ³-bench Banking is published by τ²-bench. BenchLeader re-reads the published results once a day and records the date each result was published.
- Does τ³-bench Banking count toward the BenchLeader Index?
- No. τ³-bench Banking is shown for reference but left out of the composite index.