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

AIs can outperform humans easily on short tasks, but longer ones are the true hurdle to overcome before we can deem them to ...