When a language model is optimized for reasoning, does it still show embers of autoregression? An analysis of OpenAI o1