OpenDS-World: GPT-6 Astra Leads 12 Frontier Models at 39% on Live-Web Research Tasks
A paper posted to arXiv introduces OpenDS-World, a benchmark that tasks agents with finding factual answers derived from video — then locating those answers across the live open web. Twelve frontier models were evaluated under a shared tool environment. GPT-6 Astra led all models at 39.0% accuracy across 200 tasks.
Crowd evaluators, running five participants per task across 1,000 total attempts, scored 21.1%. The best AI agent is 18 percentage points ahead of the human crowd baseline and still fails on nearly two in three tasks.
What the benchmark tests
The 200 tasks span eight domains: culture, daily life, science, economy, history, nature, society, and technology. Each task begins with a video — not a text query — and asks the agent to locate a verifiable factual answer by searching the open web. The format forces multi-modal comprehension of the video, live retrieval (web content changes over time), and search strategy management under a compute budget.
The authors are specific about where current agents break down. Tool coordination — deciding which search tools to call, in what order, and when to stop — fails on longer tasks. Budget exhaustion is the second named failure: agents spend their compute allowance before converging on an answer.
What the numbers mean
A 39% ceiling against a 21% crowd baseline produces a roughly 1.85x improvement over non-expert web search. The ceiling matters more than the ratio. OpenDS-World is contamination-resistant by design: because the answers require live web retrieval and originate from video, cached training-data knowledge cannot carry a model to the right answer — which is why these scores sit well below what the same models would post on static question-answering benchmarks.
Per-domain accuracy figures and the full model ranking are not included in the abstract. The paper does not clarify whether human evaluators worked under the same tool constraints as the agent submissions.
The OpenDS-World result fits a pattern visible across 2026’s hardest benchmarks: agents with live retrieval access consistently beat static baselines, and consistently fall short of expert human performance. The remaining gap sits in the coordination layer — selecting the right tool at the right point and managing partial results across a multi-step search horizon.