“OpenAI does not have definitive evidence that the gpt-5-thinking model could enable a novice to create severe biological harm.”