I think it's just their example chat-tuned models that are like this. Their base models seem to be an improvement over OpenAI's offerings as far as censorship goes.