There's still a lot of room for the best models to get better at coding .

Your argument rests on the "for marginal gains" part but it's really not clear that the gains are marginal in the foreseeable future.