Autism spectrum disorder (ASD) is a neurodevelopmental condition characterized by impairments in social interaction and behavior as well as structural abnormalities regarding brain development. The difference between neuroimaging-predicted brain age and the chronological age (predicted age difference; PAD) is a potential biomarker reflecting individual differences in brain developmental trajectories. However, current findings of PAD in ASD are inconsistent due to the biases introduced by different prediction models. Here, 6 classic 3-dimensional convolutional neural network (3D-CNN) models were used to assess brain age to identify consistent and reliable PAD and its interpretable features for ASD. No significant PAD differences between ASD and typically developing controls (TDC) were observed, but PAD differences in younger age and subtypes were found. Occlusion sensitivity analysis showed that default mode network (DMN) and salience network (SAN) drove the atypical brain developmental trajectories in ASD. Our findings were replicable across 6 CNN models, showing the promise of neuroimaging targets for exploring atypical developmental patterns in ASD.