OpenAI’s latest study argues that today’s top models already rival humans on real-world tasks, though it swears they won’t fully replace us.