Medical AI models must automatically adapt to differences in users, health systems, geographies, diseases, and populations to function safely and effectively.[1] This perspective proposes context switching as the defining paradigm of medical AI.[1] Context switching allows models to adapt reasoning during inference without the need for retraining.[1] Generative models thus adapt the outputs to the patient's biology, the care environment or the disease.[1] Multimodal models can reason from notes, lab results, imaging methods, and genomics, even with missing or delayed data.[1] Agent models coordinate tools and tasks according to tasks and users.[1] Context switching enables the adaptation of AI across specialties, populations, and geographies.[1] It requires advances in data design, model architecture, and evaluation frameworks.[1]