I'm sure they haven't created their own model. Probably fine-tuned it + constantly updating a database of the most occurring phrases from all LLMs.