The big powerful models think about tasks, then offload some stuff to a drastically cheaper cloud model or the model running on your hardware.