Processing perceptual sequences relies on the motor system, which is able to simulate the dynamics of the environment by developing internal representations of external events and using them to predict the incoming stimuli. Although it has previously been demonstrated that such models may incorporate predictions based on exact stimulus properties and single stimulus dimensions, it is not known whether they can also support abstract predictions pertaining to the level of stimulus categories. This issue was investigated within the present event-related functional magnetic resonance imaging study, which compared the processing of perceptual sequences of different specificity, namely those in which the sequential structure was based on the order of presentation of individual stimuli (token), and those in which such structure was defined by stimulus categories (type). The results obtained indicate a comparable engagement of the basic premotor–parietal network in processing both specific and categorical perceptual sequences. However, type sequences additionally elicited activations within the lateral prefrontal, occipital and posterior temporal regions that supported categorization in this task context. Introducing sequential deviants into token sequences activated parietotemporal and ventrolateral frontal cortices, whereas a less pronounced overall response, dominated by lateral prefrontal activation, was elicited by violations introduced into type sequences. Overall, the findings obtained suggest that, although forward models in perception may be able to incorporate expectations of lower specificity when compared to the motor domain, such processing is crucially dependent on additional contributions from lateral prefrontal as well as inferior occipital and temporal cortices that support categorization occurring in such a dynamic context.