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

OpenDS-World arrives as a direct response to a methodological problem in agent evaluation: most benchmarks rely on frozen web snapshots, which leaves open the possibility that a model's gains reflect training-data contamination rather than genuine research ability. By requiring agents to browse the live, unmodified web, locate evidence across multiple sources and file formats, and return unique factual answers, the benchmark closes that loophole. The reported gap is substantial—39.0% for GPT-6 Astra against a 21.1% human per-attempt baseline measured over 1,000 independent trials—and the shared tool environment with a common search budget means the twelve frontier models were compared under the same constraints. The paper's finding that tool coordination and long-horizon search management remain persistent failure modes across all twelve models suggests the bottleneck is not raw model capability alone but the ability to manage extended, multi-step searches. For teams building or buying web-research agents, that points to where engineering effort may be most needed.