Tax Agent Bench
Professional Work · 2026-09-23
193 held-out tax-research questions evaluated with web search, authority lookup, document fetching, information retrieval and calculation tools. GPT-5.4 judges the final answer against weighted expert rubrics, with zero credit for a failed must-pass item. Rubric credit is multiplied by 0.7 + 0.3 times independently verified citation quality. The five public and 193 private-validation questions are excluded.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Fable 5.1 | 77.6 |
| Gemini 4 Argon | 76.2 |
| Claude Opus 5 | 75.1 |
| Muse Wonder Full | 75.0 |
| DeepSeek V4.1 Flash 1017 | 74.9 |
| Qwen3.8 Max (0902) | 74.3 |
| GLM-5.3 | 73.1 |
| Muse Spark 1.3 | 72.4 |
| Gemini 4 Carbon | 71.8 |
| Muse Wonder | 71.2 |
| Grok 4.6 | 70.8 |
| Claude Fable 5 | 69.8 |
| Kimi K3 | 68.7 |
| Ember-1 | 68.0 |
| GPT-5.6 Sol | 68.0 |
| Gemini 3.8 Flash | 66.8 |
| Grok 4.7 | 65.6 |
| GPT-5.6 Terra | 65.2 |
| MiMo-V2.6-Pro | 64.9 |
| Opus 4.8 | 64.9 |
| Sonnet 4.6 | 64.9 |
| Grok 4.5 | 64.3 |
| Opus 4.7 | 63.9 |
| Hy4 Preview | 63.7 |
| Kimi K3.1 Pro | 63.7 |
| GPT-6 Astra | 63.3 |
| Mistral Large 4 Preview | 63.3 |
| MiniMax M3.1 Flash | 62.7 |
| DeepSeek-V4-Flash-0731 | 62.6 |
| DeepSeek V4.1 Flash | 62.5 |
| GPT-6.1 Sol | 62.3 |
| Sonnet 5 | 62.3 |
| Claude Haiku 5.5 | 62.2 |
| GLM-5.3 Flash | 62.0 |
| GPT-5.6 Luna | 60.8 |
| GPT-5.5 | 60.5 |
| Qwen3.8 27B | 60.0 |
| MiMo-V2.6-Flash | 59.9 |
| Step 5 Preview | 59.5 |
| GPT-6 Luna | 58.9 |