𝜏³-Banking
Agents · 2026-03-04
Τ³-Banking is a fintech customer-support benchmark in which agents retrieve policies from a large knowledge base and execute multi-step banking workflows with tools. Artificial Analysis uses the full 97-task cohort with the upstream tau2-bench v1.0.1 dataset and grader, runs five repeats per task, and reports pass@1 averaged across repeats. The task count records 97 distinct support cases, not 485 independent tasks.
Top models (higher is better)
| Model | Score |
|---|---|
| Qwen3.8-Max | 51.3 |
| Grok 4.6 | 50.7 |
| Muse Spark 1.3 | 50.5 |
| GLM-5.3 | 50.3 |
| Qwen3.8 2.4T A95B | 49.1 |
| Qwen3.8 27B | 48.0 |
| Qwen3.8 Max (0902) | 47.8 |
| GLM-5.3 Flash | 47.2 |
| Kimi K3 | 46.0 |
| Gemini 3.8 Flash | 45.8 |
| Qwen3.8-Flash-Next | 45.4 |
| Claude Opus 5 | 44.7 |
| GPT-5.6 Sol | 44.3 |
| GPT-6 Astra | 43.1 |
| Grok 4.5 | 42.1 |
| DeepSeek V4 Flash Vision Exp | 41.0 |
| GPT-5.6 Terra | 40.2 |
| DeepSeek V4 Pro 0813 | 39.6 |
| GPT-5.4 | 39.6 |
| DeepSeek-V4-Flash-0731 | 39.4 |
| GPT-5.5 | 39.0 |
| Ling 3.0 Flash Fin | 38.6 |
| Sonnet 5 | 37.3 |
| Agnes 2.5 Pro Beta | 35.7 |
| Gemini 3.7 Flash | 35.5 |
| Motif 3 | 35.3 |
| Muse Spark 1.2 | 34.8 |
| Opus 4.7 | 34.6 |
| GLM-5.2 | 34.6 |
| Sonnet 4.6 | 34.4 |
| Ling 3.0 Flash VL | 34.4 |
| Opus 4.8 | 34.2 |
| K2 Horizon 375B A23B | 34.2 |
| Gemini 3.5 Flash | 32.2 |
| K2 Horizon MoVA 36B A4B | 32.0 |
| Muse Spark 1.1 | 31.8 |
| GPT-5.6 Luna | 31.1 |
| DeepSeek-V4-Flash | 30.9 |
| DeepSeek-V4-Pro | 30.1 |
| Gemini 3.6 Flash | 29.9 |