OpenAI's Newest Reasoning Model Is Rolling Out
OpenAI is officially rolling out its latest model, o3-mini, starting today, Friday, Jan. 31. The company shared the news in a blog post on its website, just over a month after officially announcing the model during its "12 Days of OpenAI."
As with each refreshed generative AI model, o3-mini is an improvement over o1-mini—but not by as much as you might think. OpenAI says the two models perform the same in math, coding, and science, but o3-mini offers quicker answers to user queries—24% faster, in A/B testing. According to the company, testers comparing the models found o3-mini produces "more accurate and clear answers, with stronger reasoning abilities." And, with "medium reasoning effort," o3-mini matches o1 in certain reasoning and intelligence evaluations.
Like o1-mini, o3-mini is a reasoning model, a type of AI model that "thinks" through answers before responding to them. o3-mini has three different reasoning "efforts" depending on the use case: low, medium, and high. In mathematics testing, for instance, o3-mini's medium and high effort reasoning out erforms o1-mini, while high effort even outperforms o1 (the more powerful version of o1-mini). All three efforts beat o1-mini in PhD-level science questions, but o1 outperforms them all.
o3-mini replaces the o1-mini model for all users. OpenAI doesn't explicitly state why you can't use o1-mini going forward, but touts that o3-mini has higher rate limits and lower latency than the previous model.
At launch, only ChatGPT Plus, Team, and Pro users can access o3-mini. OpenAI says Enterprise users can access the model in a week. (In addition, Plus and Team users will see their daily rate limits jump from 50 messages on o1-mini to 150 messages.)
That said, free users will be able to try o3-mini in a limited capacity, either by choosing the "Reasoning" option in the message composer, or regenerating a response. OpenAI says it's the first time free users have had access to a reasoning model in ChatGPT, which comes one day after Microsoft offered o1's reasoning to Copilot users for free.
You can learn more about o3-mini in our post here. But as the model is only rolling out today, we won't know exactly how it performs until real-world testers start to use it.