
OpenAI has released its new series of “reasoning” artificial intelligence models that it claims can “think.”
The new model series, called OpenAI o1, is “designed to spend more time thinking before they respond,” the company said. The models, the first of which are available in preview in ChatGPT and through the company’s API, can “reason” through more complex tasks and problems in science, coding, and math than earlier OpenAI models. The company also announced OpenAI o1-mini, a smaller and cheaper version of the new model, which can help developers with coding tasks.
Starting on Thursday, ChatGPT Plus and Team users can access both versions of the o1 models through ChatGPT. It will have a weekly rate limit of 30 messages for the o1-preview, and 50 for o1-mini, OpenAI said. Next week, ChatGPT Enterprise and Edu users will gain access to the models. Free users of ChatGPT will eventually get access to o1-mini.
The company said it trained the models to take longer to solve problems before responding, “much like a person would.” The models were trained “to refine their thinking process, try different strategies, and recognize their mistakes,” OpenAI said.
Meanwhile, the company released evaluations of the next model update, which it said performed similarly to PhD students in testing on physics, chemistry, and biology tasks.
Unlike the current version of ChatGPT, the new model doesn’t have some of its “useful” features yet, such as browsing the web, and file and image uploads, the company said.
OpenAI also said it developed new safety training measures for o1's reasoning abilities to ensure the models follow its guidelines for safety and alignment. For example, the new model scored higher on one of OpenAI’s “hardest jailbreaking tests,” compared to its most recent model, GPT-4o.
Details about company’s new “reasoning models,” which were code-named Strawberry, were previously described by The Information and Bloomberg. Meanwhile, the AI startup is talking with investors about raising $6.5 billion at a valuation of $150 billion, Bloomberg reported. It’s also reportedly discussing raising $5 billion in debt from banks.