A symbolic matching-to-sample procedure was adopted to investigate whether humans (n=2) and baboons (n=2) discriminate more accurately the prototypes of polymorphous categories than less typical exemplars. Subjects were initially trained to discriminate between two categories of stimuli defined by the possession of any two out of three possible binary features. In transfer, prototypes, which contained all the three feature values of their categories, and novel two-out-of-three feature exemplars were presented for discrimination. Humans solved the task in a propositional way, and showed no evidence for a better performance with the prototypes than with other exemplars. By contrast, monkeys classified the prototypes more accurately than the other exemplars. The analysis of training performance showed however, that their discriminations did not involve prototypical representations of the categories, but rather depended upon feature-and exemplar-response associations. It is argued that monkeys' better performance with the prototypes rested on peak shift and/or novelty effects.