Tax Agent Bench — All-Pass
Professional Work · 2026-09-23
Percentage of the 193 held-out questions passing every rubric item, without partial credit.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Fable 5.1 | 49.2 |
| Claude Opus 5.5 | 45.6 |
| Claude Opus 5 | 45.6 |
| GLM-5.3 | 39.9 |
| Muse Spark 1.3 | 36.8 |
| Grok 4.7 | 34.2 |
| Kimi K3 | 33.7 |
| Grok 4.6 | 33.2 |
| Hy4 Preview | 32.6 |
| MiMo-V2.6-Pro | 32.6 |
| Gemini 3.8 Flash | 32.1 |
| GPT-5.6 Sol | 31.1 |
| MiMo-V2.6-Flash | 30.6 |
| Sonnet 5 | 30.1 |
| GPT-5.6 Terra | 30.1 |
| DeepSeek V4 Pro 0813 | 28.5 |
| DeepSeek V4.1 Flash | 27.5 |
| GPT-5.6 Luna | 24.9 |
| Gemini 3.7 Flash | 22.8 |
| GPT-5.5 | 21.8 |
| MiniMax M3 | 21.2 |
| GPT-6 Astra | 20.7 |
| GPT-6 Luna | 20.2 |
| Inkling | 17.6 |
| GPT-6 Sol | 15.0 |
| Mercury 2.5 | 3.6 |