Great question. Nobody knows.

The cool thing about models designed for cheaper inference is that post-training (which often requires inference as a part of the feedback loop) becomes cheaper as well, so almost certainly cheaper than models that are less optimized.