Found first: a primary source the press has not covered yet.
A paper submitted to arXiv on October 19, 2026 introduces OpenDS-World, a benchmark that tests AI agents on the live, unmodified open web, and reports that the top model, GPT-6 Astra, scores 39.0% on 200 research tasks while the human per-attempt baseline sits at 21.1% across 1,000 independent trials. The paper is Deep Search in Real-World Web Environments.
What the source says
OpenDS-World requires agents to autonomously browse the live web, locate evidence across multiple sources and file formats, and return unique factual answers. Twelve frontier models were evaluated in a shared tool environment with a common search budget. GPT-6 Astra led at 39.0% accuracy. The human baseline of 21.1% was measured across 1,000 independent per-attempt trials. The 25-author team analyzed tool usage patterns, search costs, termination behavior, and failure recovery across all twelve models.
Why it matters
Most existing benchmarks use frozen web snapshots, making it possible to attribute AI gains to training data contamination. OpenDS-World runs agents on the live web, removing that objection. On this benchmark, the top AI model scores 17.9 percentage points above the human per-attempt baseline. The paper identifies tool coordination and long-horizon search management as the persistent failure modes across all twelve models.