[m101]

Perhaps the smaller model used in o1 is over trained on arxiv and code relative to 4o (or undertrained on legal text)