Tax Agent Bench — Fact-Pattern Analysis
Professional Work · 2026-09-23
Weighted rubric and citation credit on 43 fact-pattern tax questions, using the parent evaluation protocol.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Fable 5.1 | 79.3 |
| Claude Opus 5 | 77.3 |
| Muse Spark 1.3 | 75.1 |
| Claude Opus 5.5 | 72.9 |
| Gemini 3.8 Flash | 68.0 |
| GPT-6 Astra | 66.0 |
| MiMo-V2.6-Flash | 66.0 |
| Grok 4.6 | 66.0 |
| Grok 4.7 | 64.3 |
| GPT-5.6 Terra | 63.3 |
| Kimi K3 | 62.2 |
| DeepSeek V4.1 Flash | 60.6 |
| GPT-5.6 Sol | 60.4 |
| GPT-6 Luna | 60.2 |
| Sonnet 5 | 59.9 |
| DeepSeek V4 Pro 0813 | 59.7 |
| GPT-5.6 Luna | 59.6 |
| GPT-5.5 | 59.2 |
| MiMo-V2.6-Pro | 58.4 |
| GLM-5.3 | 58.2 |
| Hy4 Preview | 55.8 |
| GPT-6 Sol | 49.9 |
| Gemini 3.7 Flash | 49.3 |
| Inkling | 44.1 |
| MiniMax M3 | 33.8 |
| Mercury 2.5 | 11.7 |