Epiq LAER CorporateBench
Professional Work · 2026-08-27
Epiq AI Labs' CorporateBench has an agent work through the email archives of four synthetic companies, from 353 to 232,692 documents, to reconstruct employees' reporting and task histories, label documents by topic and answer questions joining facts across messages. The Harbor release has 128 scored tasks over four companies and five capabilities, graded deterministically against each company's knowledge graph with exact match or set and macro F1. The headline averages the 20 company-by-capability cells equally.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-6 Astra | 82.0 |
| GPT-6 Sol | 79.0 |
| Gemini 3.8 Flash | 77.6 |
| Claude Opus 5.5 | 75.4 |
| DeepSeek V4.1 Flash | 75.3 |
| DeepSeek V4.1 Flash 1017 | 75.2 |
| Claude Opus 5 | 74.0 |
| Muse Spark Open | 73.0 |
| Kimi K3 | 72.3 |
| GPT-6 Luna | 71.7 |