The model can spend more “mental energy” on the decision because it doesn’t have to spend any on phrasing the output.