BrowseComp
BrowseComp is an OpenAI benchmark of 1,266 hard-to-find facts that tests whether an agent can keep searching until it finds a short, checkable answer.
What it means
OpenAI released BrowseComp in April 2025. Each of its 1,266 questions has a short answer that is easy to grade. The questions are built so the answer is hard to find but simple to verify once found. An agent has to search, follow links across several sites, and combine clues before it can answer.
At launch, GPT-4o with browsing scored 1.9%, and OpenAI's Deep Research scored 51.5%. Newer results exist, but many come from vendors and are not always comparable. BrowseComp measures persistence in search, not how well a tool writes a report. Hyperresearch reports 89% for its Answers product in its own runs. That is a search result, not a deep research score. The short answers make grading cheap, which is why many teams use it as a stress test for search agents.
Continue this reading path
Cite this entry
https://bestdeepresearchai.com/glossary/browsecomp/