Why Discriminative Models Often Win Asymptotically

A classifier was trained last quarter, and now only the class prevalence has shifted while the conditional shape of features given class appears stable. Which side, generative or discriminative, is easier to adjust quickly?