← All benchmarks

HyperBrowseComp

Search · 2026-10-02

MBZUAI-led benchmark of 423 manually authored, human-validated web-browsing questions in 13 languages whose concise answers require finding obscure evidence across videos, scanned documents, PDFs, images, audio, maps and tables; questions that offline models could answer were filtered out. The score is the share of questions answered correctly in one run, judged by the answering model in a fresh context against the reference answers, with execution failures counted as incorrect. Each result pairs a model with a retrieval harness: a 25-step ReAct agent using the provider's built-in search or shared Exa search and fetch tools, or the OWL multi-agent workforce.

Top models (higher is better)