Cognitive flexibility is the ability to adjust internal representations and behavior when the environment, task, or rules change. It is a core executive function, yet most computational accounts still focus on only one part of the problem, modeling local recurrent dynamics, learning mechanisms, or large-scale network behavior in isolation. In this article, we review behavioral and neural datasets commonly used to study flexibility, survey three influential model families—recurrent neural networks, reinforcement learning and meta-learning, and large-scale brain network models—and compare which aspects of flexibility each captures. Building on this, we outline a hybrid multiscale perspective in which local recurrent circuits, learning mechanisms, and whole-brain dynamics are treated as coupled components of a single system. We argue that such multiscale models can better organize existing results and guide future theoretical and empirical work.