GPT-6 Astra tops OpenDS-World deep-search benchmark at 39%

The best-performing agent in a new live-web research benchmark got 39.0% of the tasks right. Humans got 21.1%.

Those two numbers come from a paper on arXiv introducing OpenDS-World, a 200-task benchmark that asks agents to start from a video and hunt across the open web for a factual answer. The authors evaluated 12 frontier models under one tool environment. The top scorer, in the paper's own wording: "GPT-6 Astra (OpenAI, 2026b), achieves only 39.0% accuracy across all 200 tasks." Human evaluation ran five participants per task across 1,000 attempts total and landed at 21.1%.

Tasks span "eight domains: culture, daily life, science, economy, history, nature, society, and technology."

The authors are blunt that even the leader is struggling. They describe "persistent challenges in tool coordination and long-horizon search management," and name "budget exhaustion, incorrect answers at termination, and costly tool failures as key challenges for deep search." Those shortfalls, they write, "reflect limitations in current agents' information-source selection, tool coordination, search strategy adjustment, and long-horizon execution management."

The abstract omits per-domain accuracy, the full list of 12 models tested, and whether the human participants worked under the same tool environment as the agents. The 21.1% figure is a crowd-attempts baseline, not an expert-researcher one.