I think the point here is not that it does compaction (which Codex also already does) - but that the model was trained with examples of the Codex compaction, so

I think the point here is not that it does compaction (which Codex also already does) - but that the model was trained with examples of the Codex compaction, so it should perform better when compaction has taken place (a common source for drops in performance for earlier models).