I was talking about their sonar API and what you get when you go to perplexity.ai and search by default -- it's llama 3.3 70B hosted with cerebras. Which is fast, but way way behind the times -- it's more likely to hallucinate than models from basically anyone from the competition, including quick answer.

You're right though it seems that they don't just use bing + their own crawler, but who else they use seems to be community speculation.