This longitudinal study made the first attempt to employ Latent Growth Curve Modeling to analyze the development of L2 speaking accuracy and fluency through online scaffolding as well as the dynamic relationship between L2 speaking performance and self-efficacy. From the perspective of Complex Dynamic Systems Theory, it tracked the development of 45 Chinese undergraduates' English-speaking accuracy, fluency, self-efficacy for accuracy (SEA) and self-efficacy for fluency (SEF) over one semester of online teaching (six observations). Results show that speaking accuracy, SEA and SEF all improved significantly, but speaking fluency did not; these four variables all developed in non-linear trajectories, and the greatest growth of accuracy, SEA and SEF all took place at Time 2; there existed significant individual differences in the initial levels of fluency, SEA and SEF, and in the change rates of SEA; a higher initial level of accuracy was related to a greater increase in SEA and a greater decrease in growth rates with time. These findings provide evidence for non-linearity, variability and inter-individual differences in the development of L2 speaking and self-efficacy through online scaffolding, and partly confirm the dynamic relations between self-efficacy and L2 performance. Pedagogical implications for online scaffolding are also discussed.