Tax Agent Bench — Controversy & Precedence
Professional Work · 2026-09-23
Weighted rubric and citation credit on 30 tax controversy-and-precedence questions, using the parent evaluation protocol.
Top models (higher is better)
| Model | Score |
|---|---|
| Muse Wonder Full | 76.1 |
| Qwen3.8 Max (0902) | 73.5 |
| DeepSeek V4.1 Flash 1017 | 73.2 |
| Muse Wonder | 71.4 |
| GLM-5.3 | 71.1 |
| Claude Fable 5.1 | 70.6 |
| Claude Opus 5 | 70.0 |
| Kimi K3 | 68.8 |
| Muse Spark 1.3 | 68.7 |
| GPT-5.6 Sol | 68.5 |
| Claude Fable 5 | 68.4 |
| Grok 4.6 | 68.1 |
| Gemini 4 Carbon | 66.6 |
| MiMo-V2.6-Pro | 65.7 |
| GPT-5.6 Terra | 64.8 |
| GPT-6.1 Sol | 64.8 |
| Gemini 4 Argon | 63.6 |
| Grok 4.7 | 62.7 |
| GPT-5.6 Luna | 61.2 |
| Grok 4.5 | 60.7 |
| Ember-1 | 59.2 |
| Hy4 Preview | 59.1 |
| Gemini 3.8 Flash | 59.0 |
| Opus 4.8 | 58.7 |
| GPT-6 Astra | 58.6 |
| Kimi K3.1 Pro | 58.0 |
| Sonnet 4.6 | 57.9 |
| Opus 4.7 | 57.6 |
| Sonnet 5 | 57.4 |
| MiMo-V2.6-Flash | 55.7 |
| GPT-5.5 | 54.9 |
| GPT-6 Luna | 54.4 |
| MiniMax M3.1 Flash | 54.3 |
| Claude Haiku 5.5 | 54.1 |
| DeepSeek V4.1 Flash | 52.8 |
| MiniMax M3 | 52.4 |
| GLM-5.3 Flash | 52.4 |
| Step 5 Preview | 51.9 |
| DeepSeek V4 Pro 0813 | 51.5 |
| GPT-6 Sol | 50.7 |