You're reading it right. They have developed a system that automatically decides which model is sufficient, depending on your inputs, saving you costs even within one conversation stream.

The OpenAI-compatible API allows you to talk to the router like a regular GPT model.