An AI model is a computational system trained on lots of data to do a task, like writing text, recognizing images or transcribing speech. Training adjusts its parameters: numbers that decide what it answers. The app you chat in is not the model: it is the window, and the model works behind it.
A question from the app. Pick an answer.
Your app works well with a small model. What changes with a bigger one?
More parameters need more computing per token, which usually shows in the rate. Doing better on your task is not guaranteed: you measure it with your cases.
It is usually the other way around: it does more computing per token, and that takes time. Speed also depends on the hardware and how it is served.
It makes things up because it predicts what is likely, not for lack of size. A big one may invent less, but what matters still gets checked.
ChatGPT, for example, is an app. Behind it there are one or more models, and the company can switch them without renaming the app. The same goes for your bank's assistant or a WhatsApp chatbot: many of them use a model made by another company.
That is why the same app can answer differently from one month to the next. And that is why installing an app does not always install a model: many send your message to a model in the cloud and only show you the answer.
There are models for text, for images, for speech and others that combine several things. Within each type, every company trains its own and releases new versions often. Many also offer the same model in different sizes: a big one for hard tasks and a small one for fast, cheap answers.
A bigger model usually has more parameters and usually costs more per token, but that alone does not guarantee better answers for your case. With made-up rates: if the small one costs 1 USD per million tokens and the big one 8 USD, sorting 10,000 customer complaints with the big one costs eight times more. If the small one sorts them well, you overpaid.
The practical move is to test with real examples of your task, in your users' language. A model that works great in English may do worse in Spanish or Portuguese, and size will not tell you that.
"The newest model knows what happened yesterday." Not on its own: it learned up to a cutoff date and, for recent events, it needs to search or be given the information. "A model is a robot." A robot can use a model, but the model is the computation, not the body. "All AI models are chatbots." Many do not chat: they filter spam, recommend shows or read X-rays.