BrowseComp

BrowseComp questions are built with an inverted design: trainers start from a known fact and write a question whose answer does not appear on first-page search results and that another person cannot solve within ten minutes. Answers are short and graded by a model checking semantic equivalence against the reference. The benchmark is essentially unsolvable without tools (GPT-4o scored 1.9% with browsing at launch, and human trainers solved only 29.2% within a two-hour limit), so all rows here are with browsing and tools enabled, single-agent configurations only. Multi-agent harnesses score higher and are excluded to keep rows comparable; there is no single official leaderboard, so scores come from vendor model cards compiled by aggregators.