Tax Agent Bench — Citation-Adjusted Controversy & Precedence All-Pass
Professional Work · 2026-10-08
Citation-adjusted all-pass credit on 30 held-out controversy-and-precedence questions, under the October 8 headline scoring protocol.
Top models (higher is better)
| Model | Score |
|---|---|
| Muse Spark 1.3 | 36.7 |
| Claude Fable 5.1 | 26.7 |
| Muse Wonder Full | 26.7 |
| DeepSeek V4.1 Flash 1017 | 23.3 |
| Claude Opus 5 | 23.3 |
| Gemini 4 Argon | 20.0 |
| GPT-5.6 Sol | 20.0 |
| Muse Wonder | 20.0 |
| GLM-5.3 | 19.0 |
| GPT-5.6 Terra | 16.7 |
| Grok 4.6 | 16.7 |
| Claude Fable 5 | 13.3 |
| GPT-6 Astra | 13.3 |
| Grok 4.7 | 13.3 |
| Kimi K3 | 13.3 |
| MiMo-V2.6-Pro | 13.3 |
| Step 5 Preview | 13.3 |
| Opus 4.8 | 13.1 |
| Muse Spark Open | 10.3 |
| Sonnet 4.6 | 10.0 |
| Sonnet 5 | 10.0 |
| Ember-1 | 10.0 |
| GLM-5.3 Flash | 10.0 |
| GPT-5.5 | 10.0 |
| GPT-5.6 Luna | 10.0 |
| Grok 4.5 | 10.0 |
| Inkling | 9.7 |
| Muse Spark 1.2 | 7.0 |
| Claude Haiku 5.5 | 6.7 |
| Gemini 3.7 Flash | 6.7 |
| Gemini 3.8 Flash | 6.7 |
| GPT-6.1 Sol | 6.7 |
| GPT-6 Luna | 6.7 |
| GPT-6 Sol | 6.7 |
| Hy4 Preview | 6.7 |
| Inkling-Small | 6.7 |
| MiniMax M3.1 Flash | 6.7 |
| Qwen3.8 27B | 6.7 |
| MiniMax M3 | 6.5 |
| DeepSeek V4 Pro 0813 | 5.7 |