OpenAI's latest reasoning models, o3 and o4-mini, hallucinate more frequently than the company's previous AI systems, according to both internal testing and third-party research. On OpenAI's PersonQA benchmark, o3 hallucinated 33% of the time -- double the ra…