“All GPT-5 models incorporate a safety training approach called 'safe-completions' to prevent the generation of disallowed content.”