Efficiency matters too. If a model is smarter so it solves the same task with fewer tokens, that matters more than $/Mtok