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