DiligenceBench
Professional Work · 2026-07-17
DiligenceBench evaluates end-to-end research agents on 150 open-ended equity-research tasks covering 143 US public companies across ten sectors. Agents must locate filings, extract disclosures, calculate from evidence, and deliver analyst judgments; GPT-5.5 grades 5,428 task-specific weighted criteria spanning factual accuracy, analytical reasoning, and risk awareness.
Top models (higher is better)
| Model | Score |
|---|---|
| Muse Spark 1.1 | 57.4 |
| GLM-5.2 | 51.0 |
| Sonnet 4.6 | 51.0 |
| GPT-5.6 Sol | 50.0 |
| GPT-5.6 Terra | 49.1 |
| GLM-5.1 | 48.1 |
| GPT-5.6 Luna | 48.1 |
| MiniMax M3 | 47.7 |
| GPT-5.5 | 47.5 |
| Opus 4.8 | 47.1 |
| Kimi K2.6 | 43.5 |
| Gemini 3.5 Flash | 41.3 |
| Haiku 4.5 | 37.6 |
| GPT-5.4 Mini | 33.8 |
| Inkling | 32.8 |
| Gemini 3.1 Pro Preview | 30.1 |