τ²-Bench Banking (AA)
Artificial Analysis' τ²-Bench banking-domain run: a support agent using tools with a simulated customer.
As of 19 Sept 2026, Qwen3 8 leads τ²-Bench Banking (AA) on BenchLeader with 51.3%, ahead of Grok 4.6 at 50.7%, across 204 model configurations with a published result.
- Published by
- Artificial Analysis
- Category
- Agents & tools
- Index weight
- 1.0
- Models
- 204
- Data as of
- 19 Sept 2026
Source: Artificial Analysis (artificialanalysis.ai). Data taken from the public leaderboard.
What the test looks like
The banking domain of τ²-bench: the model acts as a bank's support agent, follows policy, uses tools, and coordinates with a simulated customer who also acts.
How it is scored
Percent of conversations ending in the correct final state, run by Artificial Analysis.
What to keep in mind
One domain, with a simulated customer who is itself a model. Counts toward the agentic category since index v1.2, alongside the telecom run.
- 1Qwen3 8 (max)51.3%
- 2Grok 4.6 (high)50.7%
- 3Muse Spark 1.3 (max)50.5%
- 4GLM-5.3 (max)50.3%
- 5Qwen3.8 2.4T A95B49.1%
- 6Qwen3.8 27B (xhigh)48.0%
- 7Agnes 3.0 Flash47.6%
- 8Qwen3.8 27B (medium)47.4%
- 9Claude Fable 5.1 (thinking)47.2%
- 10Muse Spark 1.3 (xhigh)47.2%
- 11GLM-5.3-Flash47.2%
- 12Kimi K3 (max)46.0%
- 13Claude Fable 5.1 (xhigh)45.8%
- 14Gemini 3.8 Flash (medium)45.8%
- 15Qwen3.8-Flash-Next45.4%
204 of 204
| # | |||
|---|---|---|---|
| 1 |  Qwen3 8maxAlibaba | 51.3% | 66.5 |
| 2 |  Grok 4.6highxAI | 50.7% | 64.5 |
| 3 |  Muse Spark 1.3maxMeta | 50.5% | 69.3 |
| 4 |  GLM-5.3maxZhipu AIopen ↗ | 50.3% | 65.8 |
| 5 | Qwen3.8 2.4T A95BAlibabaopen ↗ | 49.1% | 65.2 |
| 6 | Qwen3.8 27BxhighAlibabaopen ↗ | 48.0% | 59.2 |
| 7 |  Agnes 3.0 FlashSapiens AI | 47.6% | 62.3 |
| 8 | Qwen3.8 27BmediumAlibabaopen ↗ | 47.4% | 54.6 |
| 9 |  Claude Fable 5.1thinkingAnthropic | 47.2% | 71.3 |
| 10 | Muse Spark 1.3xhighMeta | 47.2% | 68.0 |
| 11 | GLM-5.3-FlashZhipu AIopen ↗ | 47.2% | 63.9 |
| 12 |  Kimi K3maxMoonshot AIopen ↗ | 46.0% | 67.2 |
| 13 | Claude Fable 5.1xhighAnthropic | 45.8% | 71.6 |
| 14 |  Gemini 3.8 FlashmediumGoogle | 45.8% | 64.9 |
| 15 | Qwen3.8-Flash-NextAlibabaopen ↗ | 45.4% | 61.6 |
| 16 | Gemini 3.8 FlashhighGoogle | 45.0% | 64.5 |
| 17 | Claude Opus 5highAnthropic | 44.7% | 70.2 |
| 18 |  GPT-5.6 SolmaxOpenAI | 44.3% | 68.8 |
| 19 | Grok 4.6mediumxAI | 44.3% | 66.2 |
| 20 | Claude Opus 5xhighAnthropic | 43.3% | 70.2 |
| 21 | Grok 4.6xhighxAI | 43.3% | 64.6 |
| 22 | Claude Fable 5.1highAnthropic | 43.1% | 72.0 |
| 23 | GPT-6 AstraxhighOpenAI | 43.1% | 71.1 |
| 24 | Claude Opus 5maxAnthropic | 42.1% | 69.9 |
| 25 | Grok 4.5highxAI | 42.1% | 61.0 |
| 26 | Kimi K3lowMoonshot AIopen ↗ | 41.6% | 59.6 |
| 27 | GPT-6 AstramaxOpenAI | 41.4% | 71.8 |
| 28 | Claude Fable 5.1mediumAnthropic | 41.0% | 69.6 |
| 29 |  Deepseek v4 Flash VisionmaxDeepSeekopen | 41.0% | 59.8 |
| 30 | GPT-5.6 TerramaxOpenAI | 40.2% | 65.0 |
| 31 | GPT-6 AstrahighOpenAI | 40.0% | 71.9 |
| 32 | GPT-5.4xhighOpenAI | 39.6% | 65.4 |
| 33 | DeepSeek V4 PromaxDeepSeekopen ↗ | 39.6% | 64.2 |
| 34 | DeepSeek V4 FlashmaxDeepSeekopen ↗ | 39.4% | 62.1 |
| 35 | GPT-5.5xhighOpenAI | 39.0% | 67.6 |
| 36 | Claude Fable 5.1lowAnthropic | 39.0% | 67.5 |
| 37 | Claude Opus 5mediumAnthropic | 38.6% | 67.3 |
| 38 | Ling 3.0 Flash FinUnknownopen | 38.6% | 55.4 |
| 39 | Claude Fable 5thinkingAnthropic | 38.1% | 70.9 |
| 40 | GPT-5.6 SolxhighOpenAI | 38.1% | 68.2 |
| 41 | Grok 4.6lowxAI | 38.1% | 61.0 |
| 42 | Claude Sonnet 5maxAnthropic | 37.3% | 60.5 |
| 43 | GPT-5.6 SolhighOpenAI | 36.7% | 68.0 |
| 44 | GPT-5.5highOpenAI | 36.7% | 67.0 |
| 45 | GPT-5.6 SolmediumOpenAI | 36.5% | 66.1 |
| 46 | Agnes 2.5 Pro BetaSapiens AI | 35.7% | 60.8 |
| 47 | GPT-6 AstramediumOpenAI | 35.5% | 69.8 |
| 48 | Gemini 3.7 FlashmediumGoogle | 35.5% | 64.7 |
| 49 |  Motif 3Motif Technologiesopen | 35.3% | 59.2 |
| 50 | Muse Spark 1.2xhighMeta | 34.9% | 64.1 |
| 51 | Claude Opus 4.7maxAnthropic | 34.6% | 64.3 |
| 52 | GLM-5.2maxZhipu AIopen ↗ | 34.6% | 63.9 |
| 53 | Claude Sonnet 4.6maxAnthropic | 34.4% | 57.6 |
| 54 |  Ling 3.0 Flash VLAnt Groupopen ↗ | 34.4% | 56.7 |
| 55 | Claude Opus 4.8maxAnthropic | 34.2% | 64.2 |
| 56 |  K2 Horizon 375B A23BMBZUAIopen | 34.2% | 57.6 |
| 57 | Gemini 3.8 FlashlowGoogle | 33.2% | 60.7 |
| 58 | Gemini 3.7 FlashhighGoogle | 32.8% | 64.6 |
| 59 | Gemini 3.5 FlashhighGoogle | 32.2% | 63.6 |
| 60 | Qwen3.8 27BlowAlibabaopen ↗ | 32.2% | 53.4 |
Cite as: BenchLeader, “τ²-Bench Banking (AA) leaderboard”, https://www.benchleader.com/benchmarks/aa_tau_banking, data as of 19 Sept 2026.
τ²-Bench Banking (AA): questions
- What does τ²-Bench Banking (AA) measure?
- The banking domain of τ²-bench: the model acts as a bank's support agent, follows policy, uses tools, and coordinates with a simulated customer who also acts. Scores are reported in percent of tasks solved; higher is better.
- Which AI model leads τ²-Bench Banking (AA)?
- Qwen3 8 leads τ²-Bench Banking (AA) with 51.3% as of 19 Sept 2026, ahead of Grok 4.6 at 50.7%.
- How many models have τ²-Bench Banking (AA) results?
- 204 model configurations have a τ²-Bench Banking (AA) result on BenchLeader, all taken from Artificial Analysis.
- Who runs τ²-Bench Banking (AA) and how often is it updated?
- τ²-Bench Banking (AA) is published by Artificial Analysis. BenchLeader re-reads the published results every morning and records the date each result was published.
- Does τ²-Bench Banking (AA) count toward the BenchLeader Index?
- Yes. τ²-Bench Banking (AA) contributes to the agents & tools category of the BenchLeader Index, normalised so that 50 is the average of the evaluated models.