Tax Agent Bench — Citation-Adjusted All-Pass
Professional Work · 2026-10-08
October 8, 2026 headline scoring on 193 held-out tax-research questions: each task earns credit only when every expert rubric check passes, multiplied by 0.7 + 0.3 times verified citation quality. Distinct from both weighted partial credit and the earlier unadjusted all-pass rate.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Fable 5.1 | 49.2 |
| Claude Opus 5 | 45.5 |
| Gemini 4 Argon | 44.6 |
| DeepSeek V4.1 Flash 1017 | 44.4 |
| Muse Wonder Full | 43.0 |
| Muse Spark 1.3 | 42.4 |
| GLM-5.3 | 39.7 |
| Claude Fable 5 | 38.3 |
| Muse Wonder | 36.3 |
| Grok 4.7 | 34.2 |
| Kimi K3 | 33.4 |
| Grok 4.6 | 33.1 |
| Muse Spark 1.2 | 32.8 |
| Hy4 Preview | 32.6 |
| MiMo-V2.6-Pro | 32.4 |
| MiniMax M3.1 Flash | 32.1 |
| Gemini 3.8 Flash | 32.1 |
| Claude Haiku 5.5 | 31.9 |
| GLM-5.3 Flash | 31.6 |
| GPT-5.6 Sol | 31.1 |
| Ember-1 | 31.1 |
| Qwen3.8 27B | 30.9 |
| Mistral Large 4 Preview | 30.4 |
| GPT-5.6 Terra | 30.0 |
| Opus 4.8 | 30.0 |
| Sonnet 5 | 29.4 |
| MiMo-V2.6-Flash | 28.5 |
| DeepSeek-V4-Flash-0731 | 28.1 |
| Step 5 Preview | 27.3 |
| DeepSeek V4 Pro 0813 | 26.8 |
| DeepSeek V4.1 Flash | 26.6 |
| Muse Spark Open | 26.3 |
| Sonnet 4.6 | 25.8 |
| Grok 4.5 | 24.9 |
| GPT-5.6 Luna | 24.8 |
| Opus 4.7 | 23.7 |
| Gemini 3.7 Flash | 22.8 |
| GPT-5.5 | 21.8 |
| GPT-6.1 Sol | 21.8 |
| Gemini 3.5 Flash | 21.7 |