People say model when they mean the app, and app when they mean the model. The distinction matters, because most of what you experience day to day comes from the product, and most of the headlines are about the model.
The model is the part that was trained: a file full of learned numbers that turns input into output. Everything else, the chat window, the saved history, the connected calendar, is software built around it.
The model is a frozen artifact
After training, the model’s numbers are locked in. It has no clock, no account for you, and no idea which app it is running inside. Send it the same input twice in a clean room and you get roughly the same behavior, because nothing about talking to you changed it.
The freeze is actually useful. Because the model does not change, it can be tested and trusted to behave tomorrow the way it behaved today. That is why providers give models version names: so the products built on top can choose exactly when to move to a new engine instead of having the ground shift under them.
The product does the remembering
When an assistant knows your name across sessions, that is the product saving a note and showing it to the model again next time. When it checks your calendar, that is the product granting a tool. The model supplies the intelligence, and the product supplies everything personal.
A good comparison is an engine and a car. The same engine can power a delivery van or a sports car, and most of what you notice as the driver comes from the car: the seats, the steering, the safety features. It works the same way here. The model sets the ceiling on how good the product can be, and the product decides how much of that ceiling you actually experience.
Why you should care
When something feels smart, thank both layers. When something feels off, ask which layer failed. A model can write a weak paragraph, but forgetting your name, losing a file, or emailing the wrong person is product behavior. Knowing where the line sits makes you much better at judging AI tools.
This also explains why the same app can feel instant on easy questions and slower on hard ones. Products pick engines per task, using a small quick model to summarize a page and a bigger one for a hard question. That kind of routing is a sign of a well-built product doing its job.