Computational thinking (CT), the ability to devise computational solutions for real-life problems, has received growing attention from both educators and researchers. To better improve university students' CT competence, collaborative programming is regarded as an effective learning approach. However, how novice programmers develop CT competence through collaborative problem solving remains unclear. This study adopted an innovative approach, quantitative ethnography, to analyze the collaborative programming activities of a high-performing and a low-performing team. Both the discourse analysis and epistemic network models revealed that across concepts, practices, and identity, the high-performing team exhibited CT that was systematic, whereas the CT of the low-performing team was characterized by tinkering or guess-and-check approaches. However, the low-performing group's CT development trajectory ultimately converged towards the high-performing group's. This study thus improves understanding of how novices learn CT, and it illustrates a useful method for modeling CT based in authentic problem-solving contexts.