Tax Agent Bench — Rule & Source Lookup
Professional Work · 2026-09-23
Weighted rubric and citation credit on 30 rule-and-source tax questions, using the parent evaluation protocol.
Top models (higher is better)
| Model | Score |
|---|---|
| GLM-5.3 | 81.3 |
| Claude Opus 5 | 78.0 |
| Claude Fable 5.1 | 72.9 |
| Muse Spark 1.3 | 67.7 |
| Grok 4.6 | 67.5 |
| Sonnet 5 | 61.5 |
| Gemini 3.8 Flash | 60.2 |
| MiMo-V2.6-Pro | 59.8 |
| Kimi K3 | 58.4 |
| Gemini 3.7 Flash | 57.9 |
| MiniMax M3 | 57.9 |
| Hy4 Preview | 57.5 |
| DeepSeek V4.1 Flash | 56.9 |
| MiMo-V2.6-Flash | 56.3 |
| Claude Opus 5.5 | 56.2 |
| GPT-5.6 Luna | 54.3 |
| GPT-5.6 Sol | 53.8 |
| GPT-5.6 Terra | 53.3 |
| Grok 4.7 | 52.2 |
| DeepSeek V4 Pro 0813 | 49.8 |
| GPT-6 Astra | 49.4 |
| GPT-6 Luna | 45.9 |
| GPT-5.5 | 44.2 |
| GPT-6 Sol | 39.8 |
| Inkling | 31.3 |
| Mercury 2.5 | 8.3 |