It makes sense to optimize poly counts and textures during development. When they're building the first models they have no clue on what exactly the engine can do in terms of scene complexity, so they have to make guesses and hope that it'll all turn out right. It's all right if first estimations turn out to be wrong.
But as I've said before it also makes sense to do several passes on all artwork during development. So first get everything in the game, to be able to test it for performance, looks, gameplay and all; and then use all that you've learned to rework everything as good as you can. GTA4 for example had at least 3 different versions of Niko since the first screenshots from last May.