I'm not entirely convinced the new models aren't just a rollback to a clean training data state after the current model has been released for a while.