It uses hidden models, mostly Claude and open ai models. They try to check the prompt and route to the model that is capable of it, thus saving money.

I think this is it ^^^ Routing and optimizations :)