Subdivision models naturally have more polygons than realtime models because of holding edges and good edge flow overall. However, it’s important to realize that a really well-made subdivision model often has fewer polygons than a non-subdividable model.
The following two images show a model of a Dimplex grill. On the left is the original model an artist created. He tried to make the model so it didn’t need subdivision when used as a background or mid-range object. But the model had no holding edges, so it couldn’t be subdivided for close-ups. The model on the right, however, was made especially for subdivision, with fewer segments along the side. With no subdivision the model works for background shots, at Level 1 subdivision for mid-range shots, and Level 2-3 for closeups. The subdividable model, at the base level, has fewer polygons than the original model.
Another example is this skull model. The original model (left) has a high number of polys. The same model can be represented with a much lower poly count (middle) with the option to subdivide (right). This gives the customer more options while keeping the base poly count fairly low for background use.