Sometimes expensive models are cheaper on easier tasks because they use fewer tokens, too.