The authors introduce a deep learning framework to reproduce sequences of response times and use it to provide evidence for a stability–flexibility trade-off underlying task-switching costs.