Nimble Web Context
Search
Nimble's Web Context benchmark tests open-web research agents on go-to-market, finance, and social-media work: enriching a company or creator, finding every entity that matches a set of criteria, and writing a research report. Nimble scores 90 tasks, 30 per vertical drawn from a 300-task pool, with agents limited to web search and page extraction. LLM judges verify each reported fact against its source pages, and the score is per-task F1 of fact precision and recall averaged over the 90 tasks.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-5.6 Sol | 80.0 |
| Qwen3.8-Max | 77.3 |
| GPT-6 Sol | 76.5 |
| DeepSeek V4.1 Flash | 73.5 |
| GPT-6 Astra | 72.6 |
| Claude Opus 5.5 | 71.8 |
| GLM-5.3 | 69.8 |
| Kimi K3 | 69.4 |
| GPT-6 Luna | 67.9 |
| Claude Sonnet 5.5 | 67.2 |
| GLM-5.3 Flash | 64.2 |