Tax Agent Bench — Numbers & Calculations
Professional Work · 2026-09-23
Weighted rubric and citation credit on 30 quantitative tax questions, using the parent evaluation protocol.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5.5 | 88.9 |
| Claude Fable 5.1 | 88.0 |
| Claude Opus 5 | 85.2 |
| GLM-5.3 | 84.5 |
| Kimi K3 | 84.0 |
| Gemini 3.8 Flash | 83.1 |
| Muse Spark 1.3 | 81.8 |
| GPT-5.6 Terra | 81.6 |
| GPT-5.6 Sol | 81.4 |
| Hy4 Preview | 81.0 |
| Grok 4.6 | 79.8 |
| Gemini 3.7 Flash | 78.8 |
| GPT-6 Astra | 77.9 |
| GPT-5.5 | 76.8 |
| Grok 4.7 | 76.7 |
| DeepSeek V4.1 Flash | 76.6 |
| MiMo-V2.6-Pro | 75.5 |
| GPT-6 Luna | 73.8 |
| GPT-5.6 Luna | 71.8 |
| Sonnet 5 | 70.6 |
| DeepSeek V4 Pro 0813 | 70.2 |
| GPT-6 Sol | 69.5 |
| MiMo-V2.6-Flash | 62.4 |
| Inkling | 61.9 |
| MiniMax M3 | 54.9 |
| Mercury 2.5 | 31.1 |