The latent class structure of autism symptoms from the time of diagnosis to age 6 years was examined in a sample of 280 children with autism spectrum disorder. Factor mixture modeling was performed on 26 algorithm items from the Autism Diagnostic Interview-Revised at diagnosis (Time 1) and again at age 6 (Time 2). At Time 1, a "2-factor/3-class" model provided the best fit to the data. At Time 2, a "2-factor/2-class" model provided the best fit to the data. Longitudinal (repeated measures) analysis of variance showed that the "2-factor/3-class" model derived at the time of diagnosis allows for the identification of a subgroup of children (9% of sample) who exhibit notable reduction in symptom severity.