RedlineBench
Professional Work
Crosby and micro1's RedlineBench: 140 Harbor tasks in which an agent redlines commercial contracts as a .docx with tracked changes and comments over three multi-turn MSA negotiations. Attorney rubrics are graded by an LLM judge panel, and the headline is a turn-weighted mean score.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-6 Astra | 62.8 |
| Grok 4.6 | 54.7 |
| Claude Opus 5 | 50.7 |
| GPT-5.6 Sol | 50.3 |
| Ember-1 | 50.2 |
| Muse Spark Open | 49.9 |
| GLM-5.3 | 49.6 |
| Kimi K3 | 49.0 |
| DeepSeek V4 Pro 0813 | 46.9 |
| DeepSeek V4.1 Flash 1017 | 44.9 |
| Gemini 3.8 Flash | 44.5 |
| GPT-5.5 | 44.1 |
| GPT-5.6 Terra | 44.0 |
| GPT-5.6 Luna | 43.3 |
| Claude Fable 5 | 42.4 |
| Gemini 3.6 Flash | 41.6 |
| Claude Haiku 5.5 | 39.6 |
| GLM-5.2 | 39.6 |
| Gemini 3.5 Flash | 39.1 |
| Opus 4.8 | 38.2 |