DeepResearch Bench II (Mercor 130-task set)
Search
DeepResearch Bench II as Mercor runs it: 130 bilingual research briefs scored against 9,287 expert-written rubric criteria, against the original release's 132 tasks and 9,430 rubrics. The total score aggregates rubric assessments per task.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5 | 56.1 |
| Claude Fable 5.1 | 55.0 |
| Claude Opus 5.5 | 53.3 |
| Sonnet 4.6 | 53.3 |
| DeepSeek V4.1 Flash | 53.2 |
| GLM-5.3 | 52.0 |
| Claude Sonnet 5.5 | 51.7 |
| GPT-5.6 Sol | 51.5 |
| GPT-6 Astra | 50.1 |
| GPT-5.6 Luna | 49.7 |
| Grok 4.7 | 49.0 |
| GPT-5.6 Terra | 47.2 |
| Grok 4.5 | 47.1 |
| Opus 4.7 | 45.4 |
| GPT-5.5 | 45.4 |
| Opus 4.8 | 44.6 |
| Gemini 3.7 Flash | 43.4 |
| GPT-6 Luna | 42.9 |
| GPT-6 Sol | 42.9 |
| Sonnet 5 | 39.7 |
| GPT-5.4 | 38.4 |
| Gemini 3.6 Flash | 36.3 |
| Claude Haiku 5.5 | 34.3 |