BrowseComp
BrowseComp is a benchmark of questions that need several rounds of web browsing to answer.
What it means
BrowseComp is a benchmark of questions that are hard to answer without browsing. A correct answer usually needs several searches, reads of several pages, and a step that ties the facts together. Vendors publish scores on it because the test checks whether a system can find and combine scattered information from the web rather than recall a single fact.
Scores depend on the test setup. The model behind the agent, the number of steps allowed, the grader and the search tool all change the result. Hyperresearch publishes BrowseComp scores of 89% for Hyperresearch Answers, 73% for Parallel (Pro), 58% for OpenAI web search, 43% for Perplexity Sonar Pro and 12% for Exa, from its October 2026 runs. Parallel publishes its own figures, so compare vendor numbers with care and run a small test on your own queries.
Continue this reading path
Cite this entry
https://apisearchengine.com/glossary/browsecomp/