“The parallel processing of reasoning steps in the o1 system makes it incompatible with the token-streaming output common in other language models.”