According to internal tests, newer models like o3 and o4-mini hallucinate significantly more than older versions, and OpenAI doesn't know why.