OpenAI introduces o1, its first reasoning model series

Background

The o1 series was designed to spend more time thinking before it responded, which OpenAI aimed at complex tasks in science, coding and math ^[1].

What happened

On 12 September 2024 OpenAI introduced two models. o1-preview was the early version of the model. o1-mini was a faster version, particularly effective at coding, which Microsoft said cost about 80% less than o1-preview ^[1][4]. OpenAI's system card says the o1 series is trained with large-scale reinforcement learning to reason using chain of thought, and that it can produce a long chain of thought before answering the user ^[2].

The same day Microsoft added o1-preview and o1-mini to Azure OpenAI Service, Azure AI Studio and GitHub Models, as a preview for select customers ^[4].

Key details

OpenAI reported that reasoning about its safety policies in context gave the models state-of-the-art performance on certain risk benchmarks, such as resisting known jailbreaks and avoiding illicit advice ^[2]. Under OpenAI's Preparedness Framework, both models were rated overall medium risk: medium for persuasion and CBRN (chemical, biological, radiological and nuclear) risk, and low for model autonomy and cybersecurity ^[2].

The full o1 model followed. OpenAI published an updated o1 system card on 5 December 2024, and introduced o1 for developers on 17 December 2024 ^[3][5].

Impact

The launch introduced OpenAI's first reasoning model series, which trades response time for more reasoning. It also reached developers and enterprises early through Azure, as a preview for select customers ^[4].