Tax Agent Bench — Citation-Adjusted Numbers & Calculations All-Pass
Professional Work · 2026-10-08
Citation-adjusted all-pass credit on 30 held-out quantitative questions, under the October 8 headline scoring protocol.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Fable 5.1 | 60.0 |
| Claude Opus 5 | 60.0 |
| Gemini 4 Argon | 60.0 |
| Hy4 Preview | 60.0 |
| MiniMax M3.1 Flash | 59.9 |
| Claude Haiku 5.5 | 53.3 |
| GLM-5.3 | 53.2 |
| Claude Fable 5 | 50.0 |
| Gemini 3.8 Flash | 50.0 |
| Muse Spark 1.2 | 48.7 |
| Sonnet 5 | 47.0 |
| Opus 4.8 | 46.7 |
| Gemini 3.5 Flash | 46.7 |
| GLM-5.3 Flash | 46.7 |
| GPT-5.6 Sol | 46.7 |
| Grok 4.6 | 46.7 |
| Muse Wonder Full | 46.7 |
| DeepSeek V4.1 Flash 1017 | 45.7 |
| Kimi K3 | 45.7 |
| MiMo-V2.6-Pro | 45.7 |
| Muse Spark 1.3 | 43.3 |
| DeepSeek-V4-Flash-0731 | 41.3 |
| Sonnet 4.6 | 40.0 |
| GPT-5.6 Terra | 40.0 |
| Grok 4.7 | 40.0 |
| Mistral Large 4 Preview | 40.0 |
| Muse Wonder | 40.0 |
| Gemini 3.6 Flash | 39.7 |
| Muse Spark Open | 38.7 |
| DeepSeek V4.1 Flash | 38.0 |
| Opus 4.7 | 36.7 |
| Ember-1 | 36.7 |
| Gemini 3.7 Flash | 33.3 |
| GPT-6.1 Sol | 33.3 |
| Grok 4.5 | 33.3 |
| MiMo-V2.6-Flash | 32.7 |
| DeepSeek V4 Pro 0813 | 32.3 |
| Qwen3.8 27B | 32.3 |
| GPT-5.5 | 30.0 |
| GPT-5.6 Luna | 30.0 |