Found first: a primary source the press has not covered yet.
A 25-author team spanning HKUST(GZ), Fudan University, Stanford, and Microsoft Research has published OpenDS-World, a benchmark testing AI deep-search agents on the live, unmodified open web. GPT-6 Astra reached 39.0% accuracy on 200 tasks across eight domains, while human participants scored 21.1% per attempt across 1,000 independent trials.
What the source says
OpenDS-World covers 200 tasks across eight domains: culture, daily life, science, economy, history, nature, society, and technology. Twelve frontier models were evaluated alongside human participants, all using shared tools and the same search budget. Tasks run on the live web, not archived snapshots. GPT-6 Astra led at 39.0%, followed by Claude Opus 5.5 at 38.0% and Gemini 3.8 Flash at 36.5%. The benchmark finds substantial variation in research strategies across models and identifies tool coordination and long-horizon search management as persistent failure modes.
Why it matters
Running evaluation on the live web removes the memorization confound that affects snapshot-based benchmarks. At 39.0%, the leading model already exceeds the human per-attempt rate on tasks requiring real-time open-web research, which moves the baseline for evaluating this class of agent. Tool coordination and long-horizon search management are the persistent failure modes the paper identifies, and they remain unsolved across all 12 evaluated models.