Reactive joint attention (RJA) describes shared attention on a cued target. This key ability is attenuated in autistic compared to non-autistic preschoolers with low cognitive ability, and thus trained during early intervention. We evaluated the development of RJA in matched autistic preschoolers within a randomized controlled trial of the naturalistic developmental behavioral intervention A-FFIP (intervention [n = 32] versus early intervention-as-usual [EIAU, n = 28]), which is further compared to non-autistic preschoolers (n = 52). A screen-based eye-tracking paradigm assessed RJA at baseline, after 12 months (end-of-intervention, 78% retention), and after 36 months (follow-up, 44% retention). Corresponding pupil size changes were utilized to investigate arousal as a mediator in RJA group differences. Generalized linear mixed models were applied to compare RJA likelihood between groups and assessment timepoints. Across timepoints, RJA likelihood was lower in autistic versus non-autistic preschoolers (ORs = 0.07-0.27). The A-FFIP - but not the EIAU group - showed an increase in RJA likelihood at end-of-intervention (OR = 1.52) and follow-up (OR = 2.38). Across both autistic groups, an increase in RJA likelihood after 12 months predicted improved social responsiveness at 36-months follow-up (β = -1.22). A higher baseline pupil size within trials was associated with a lower RJA likelihood (β = -0.32) and mediated the autistic group difference on RJA likelihood in a causal mediation analysis. The A-FFIP early intervention increased eye-tracking derived RJA in autistic preschoolers up to two years after end of intervention, which likely cascaded on improved social responsiveness. Arousal regulation is outlined as a promising mediating mechanism.