BrowseComp (without browsing)
Search · 2025-04-10
BrowseComp's 1,266 questions answered without web browsing, from the model's own knowledge, as in the BrowseComp paper's baselines. It measures closed-book recall of obscure facts rather than browsing, so it is excluded from scalar capability; browsing runs are scored under BrowseComp.
Benchmark source: OpenAI
Top models (higher is better)
| Model | Score | Source |
|---|---|---|
| o1 (2024-12-17) | 9.9 | BrowseComp benchmark release · OpenAI |
| GPT-4.5 | 0.9 | BrowseComp benchmark release · OpenAI |
| GPT-4o (2024-08-06) | 0.6 | BrowseComp benchmark release · OpenAI |