To investigate temperament as an early risk marker for autism spectrum disorder (ASD), we examined parent-reported temperament for high-risk (HR, n = 170) and low-risk (LR, n = 77) siblings at 8, 14, and 24 months. Diagnostic assessment was performed at 36 months. Group-based analyses showed linear risk gradients, with more atypical temperament for HR-ASD, followed by HR-Atypical, HR-Typical, and LR siblings. Temperament differed significantly between outcome groups (0.03 ≤ ηp2 ≤ 0.34). Machine learning analyses showed that, at an individual level, HR-ASD siblings could not be identified accurately, whereas HR infants without ASD could. Our results emphasize the discrepancy between group-based and individual-based predictions and suggest that while temperament does not facilitate early identification of ASD individually, it may help identify HR infants who do not develop ASD.