New audio models from OpenAI but how much can we rely on them