Several previous studies have reported atypicality in resting-state functional connectivity (FC) in autism spectrum disorder (ASD), yet the relatively small effect sizes prevent us from using these characteristics for diagnostic purposes. Here, canonical correlation analysis (CCA) and hierarchical clustering were used to partition the high-functioning ASD group (i.e., the ASD discovery group) into subgroups. A support vector machine (SVM) model was trained through the 10-fold strategy to predict Autism Diagnostic Observation Schedule (ADOS) scores within the ASD discovery group (r = 0.30, P < 0.001, n = 260), which was further validated in an independent sample (i.e., the ASD validation group) (r = 0.35, P = 0.031, n = 29). The neuroimage-based partition derived two subgroups representing severe versus mild autistic patients. We identified FCs that show graded changes in strength from ASD-severe, through ASD-mild, to controls, while the same pattern cannot be observed in partitions based on ADOS score. We also identified FCs that are specific for ASD-mild, similar to a partition based on ADOS score. The current study provided multiple pieces of evidence with replication to show that resting-state functional magnetic resonance imaging (rsfMRI) FCs could serve as neural biomarkers in partitioning high-functioning autistic individuals based on their symptom severity and showing advantages over traditional partition based on ADOS score. Our results also indicate a compensatory role for a frontocortical network in patients with mild ASD, indicating potential targets for future clinical treatments.