They're likely referring to using the large language models (GPT, Claude, etc.) through their primary interfaces, because you get more control by interacting with the LLM directly.