A discrepancy between first- and third-party benchmark results for OpenAI's o3 AI model is raising questions about the company's transparency and model testing practices.