τ³-Banking (Fireworks, banking_knowledge)
Agents
Fireworks' run of the 97-task banking_knowledge subset of τ³-Bench 1.0.1, in which policy-bound banking support agents use database tools and a knowledge base with a simulated customer. Fireworks runs Harbor 0.20 with Tau3LLMAgent over MCP and alltools retrieval, four official seeded trials per task, a GPT-5.2 user simulator and judge, and each model at its highest available thinking level. The score is the mean binary trial reward; this protocol is kept as a component beside Artificial Analysis's five-repeat banking run.
Top models (higher is better)
| Model | Score |
|---|---|
| Gemini 4 Argon | 58.2 |
| GLM-5.3 | 51.6 |
| Kimi K3 | 51.5 |
| Gemini 3.8 Flash | 51.4 |
| Qwen3.8-Max | 51.3 |
| Grok 4.6 | 50.7 |
| Claude Opus 5 | 49.1 |
| GPT-6 Sol | 47.0 |
| DeepSeek V4.1 Flash 1017 | 46.8 |
| GPT-6 Astra | 44.8 |
| DeepSeek V4 Pro 0813 | 42.7 |
| Ember-1 | 38.1 |
| GPT-6 Luna | 37.7 |
| Muse Spark Open | 37.2 |
| DeepSeek V4.1 Flash | 35.7 |
| GPT-5.6 Sol | 35.7 |