<p>As artificial intelligence coding assistants (AICAs) are increasingly adopted in K–12 coding education, the role of students' computational thinking (CT) levels in the learning process remains under-explored. While existing research focuses on overall outcomes, this study explores the impact of CT levels on students’ learning performance, behavioral patterns, and cognitive patterns to reveal the underlying process-level mechanisms. 78 eighth-grade students (aged 13–15) from a public secondary school participated in a four-week AICA-assisted Python coding course. According to the results of a validated CT pre-test, the top and bottom thirds of students were selected for comparative analysis, yielding a final sample of 52 students (26 high CT and 26 low CT). Students’ learning performance was assessed through pre- and post-tests, while their screen-capture videos and reflection journals were analyzed using Lag Sequential Analysis (LSA) and Epistemic Network Analysis (ENA) to examine behavioral trajectories and cognitive structure coherence. The results showed that the learning performance of the high CT group significantly outperformed the low CT group. Behaviorally, the high CT students used AICA for better code understanding, while the low CT group used AICA for immediate answer retrieval. Cognitively, the ENA of the reflection journals revealed that the high CT group demonstrated stronger self-regulatory coherence, displaying connected transitions among the planning, execution, and self-reflection phases—a structure that was absent in the low CT group. These findings underscore the profound impact of CT level on students’ interaction with and learning from AICA. Two differentiated instructional strategies are proposed: providing open-ended AICA support for high CT students to promote independent exploration, and offering structured guidance via AICA to low CT students to enhance their independent learning and self-regulated learning skills.</p>