Categorisation and such models are wayyyyyy cheaper to train and run than LLMs. And the training data is just their own price and sales history by location, nothing to scrape the whole Internet or scan unknown books for. This costs them very little at their scale. That said, their dynamic pricing would have to halve the prices to get me to go again. Because the price of a big tasty meal is nearly twice what a double bug tasty meal was 6 or 7 years ago. Beef has also nearly doubled, but the other ingredients have not.