OpenDS-World: GPT-6 Astra 39.0% Beats Humans' 21.1%

The team behind OpenDS-World spans HKUST(GZ), Fudan University, Stanford, and Microsoft Research — a 25-author collaboration. Their benchmark covers 200 tasks across eight domains: culture, daily life, science, economy, history, nature, society, and technology. All twelve frontier models and the human participants used shared tools and the same search budget, so the comparison holds conditions constant.

The benchmark finds substantial variation in research strategies across models. That variation, combined with the persistent failure modes of tool coordination and long-horizon search management, suggests the gap between the leading 39.0% and the human 21.1% per-attempt rate is not the whole story — the remaining errors cluster around how agents manage multi-step searches rather than raw knowledge. Because tasks run on the live web, results may shift as the web itself changes, a property snapshot benchmarks do not have.