Odysseys2 launches with 100 web-agent tasks based on real user requests

Researchers at Carnegie Mellon University and Browser Use released Odysseys2, a benchmark of 100 web-navigation tasks built from the distribution of real requests to Browser Use rather than browsing histories. They report that GPT-6 Astra led 15 agents, with perfect scores on 47 tasks. Their judge considers delivered files alongside browsing actions and screenshots, and gives zero credit to attempts it flags for suspected reward hacking.