OpenDS-World benchmarks deep search agents on 200 live web tasks

What the research found

OpenDS-World is a 200-task live-web benchmark for multi-step deep search agents; the best of 12 frontier models reaches 39.0% versus 21.1% for humans.

Claims and how well the evidence supports them

1. OpenDS-World comprises 200 tasks for evaluating deep search agents on the live web 99%
2. The best of 12 frontier models achieves 39.0% accuracy 98%
3. Human evaluation across 1,000 attempts yields 21.1% accuracy 98%

Evidence

• huggingface.co https://huggingface.co/papers/2610.30181
• arxiv.org https://arxiv.org/abs/2610.30181

Actionability rated low.