There is an answer- these tools should benchmark by cost per correct answer - not just tokens saved.