Tax Agent Bench — Rule & Source All-Pass
Professional Work · 2026-09-23
All-rubric-item pass rate on the 30 rule-and-source questions.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Fable 5.1 | 66.7 |
| GLM-5.3 | 60.0 |
| Claude Opus 5 | 56.7 |
| Claude Opus 5.5 | 50.0 |
| Sonnet 5 | 43.3 |
| Grok 4.6 | 43.3 |
| Kimi K3 | 43.3 |
| DeepSeek V4.1 Flash | 40.0 |
| DeepSeek V4 Pro 0813 | 40.0 |
| Gemini 3.7 Flash | 40.0 |
| Grok 4.7 | 40.0 |
| MiMo-V2.6-Flash | 40.0 |
| Muse Spark 1.3 | 40.0 |
| Gemini 3.8 Flash | 36.7 |
| GPT-5.6 Luna | 36.7 |
| Hy4 Preview | 36.7 |
| MiniMax M3 | 36.7 |
| GPT-5.6 Terra | 33.3 |
| MiMo-V2.6-Pro | 33.3 |
| GPT-5.6 Sol | 30.0 |
| GPT-5.5 | 23.3 |
| GPT-6 Astra | 23.3 |
| GPT-6 Luna | 20.0 |
| GPT-6 Sol | 16.7 |
| Inkling | 16.7 |
| Mercury 2.5 | 3.3 |