Diagnostic criteria used in autism research have undergone a shift towards the inclusion of a larger population, paralleled by increasing, but variable, estimates of autism prevalence across clinical settings and continents. A categorical diagnosis of autism spectrum disorder is now consistent with large variations in language, intelligence, comorbidity, and severity, leading to a heterogeneous sample of individuals, increasingly distant from the initial prototypical descriptions. We review the history of autism diagnosis and subtyping, and the evidence of heterogeneity in autism at the cognitive, neurological, and genetic levels. We describe two strategies to address the problem of heterogeneity: clustering, and truncated-compartmentalized enrollment strategy based on prototype recognition. The advances made using clustering methods have been modest. We present an alternative, new strategy for dissecting autism heterogeneity, emphasizing incorporation of prototypical samples in research cohorts, comparison of subgroups defined by specific ranges of values for the clinical specifiers, and retesting the generality of neurobiological results considered to be acquired from the entire autism spectrum on prototypical cohorts defined by narrow specifiers values.