A capability that allows a model to switch between fast, direct responses and slower, more deliberate reasoning depending on task complexity.