For two years the reflex was to send everything to a large model in the cloud. That still makes sense for the hardest tasks, but it is no longer the only option. A new generation of compact models runs on the phone itself — summarising a note, classifying a photo, drafting a reply — without a round trip to a server. The result is instant, works offline, and costs nothing per request.
We now treat on-device as the default and the cloud as the escalation. Simple, high-volume features stay local, which keeps user data on the device and the bill predictable; only the genuinely hard requests go to a larger model. For most apps that mix is cheaper, faster, and far easier to defend to privacy-conscious users — a better deal than routing every tap through someone else's datacentre.