OpenWorldEngineer (Mean@8)

Mercor's OpenWorldEngineer (OWE-Bench) gives an agent a long-horizon engineering or research project and judges how far it advances it: 26 expert-built tasks in science, software, machine learning and engineering, with 4 to 18 hours and the task's hardware (up to a B300) per attempt, web search over a copy of the web frozen at October 1, 2026, and a deterministic verifier that scores a frozen artifact on held-out cases the agent never sees. Each attempt is rescaled so the untouched baseline scores 0 and an expert-set ceiling (an exact answer, an engineering bound or an attainable estimate) scores 100; a task counts the mean of 8 attempts, and the score is the mean over tasks. Agents run in Mercor's open OpenWorld harness on Harbor at the effort the leaderboard lists.