ChatGPT has a new model called o1 that is trained to solve more difficult problems, analyze its answers, try different strategies and refine its thinking, OpenAI said in a blog post on Thursday.

The new model, currently split between o1-preview and o1-mini, ranks in the 89th percentile in Codeforces programming competitions, is among the top 500 students in the U.S. in the Math Olympiad, and “exceeds the accuracy of a PhD student on a benchmark of physics, biology, and chemistry problems,” according to OpenAI.

“We found that this model hallucinates less,” said Jerry Tworek, head of research at OpenAI, in an interview with The Verge. It was trained using a new optimization algorithm and a customized training dataset. While previous models aimed to mimic patterns in their training data, o1 uses reinforcement learning, which trains it through rewards and punishments.

What sets o1 apart from previous models is its ability to “think,” according to a report from The Information on Tuesday. That means the model doesn’t start spitting out answers immediately, but takes 10 to 20 seconds to put together a well-thought-out response. The o1 model, also dubbed “Strawberry” by viewers (a possible reference to the viral trend where influencers ask AIs how many “R’s” are in the word “Strawberry”), does away with the “thought chain prompt” that requires users to ask an AI additional questions to see its intermediate reasoning. Instead, the model is designed to show its reasoning by default.

Because o1 is still in preview, it has some significant limitations. Unlike GPT-4o, o1 is not connected to the Internet, cannot be used with file uploads, and has a variety of API limitations for developers. The o1-mini model differs in that it focuses on providing quick answers to STEM-related questions.

Competition in the AI ​​space is getting fiercer as all the players in Big Tech try to outdo each other and develop “agent-based” AIs that can do tasks for you. At Google I/O earlier this year, the search engine giant unveiled a more powerful version of Gemini that can communicate with you more naturally and even allow you to interrupt it mid-sentence. And at the iPhone 16 launch event earlier this week, Apple boosted the processing power of its latest handsets to be able to handle Apple Intelligence, a set of AI features for the iPhone based on OpenAI technology.

While the hype around AI has driven record highs in tech stocks over the past two years, investors appear to be growing more cautious. Nvidia, the chipmaker that designs the brains of many of the world’s leading AI data centers, saw its shares fall 10% last week. The tech world’s overall sentiment around AI may be cooling while it waits for more concrete results from services, though that hasn’t stopped OpenAI from reaching a staggering $150 billion valuation.

For ChatGPT Plus and Team users, the o1 preview model is rolling out now. ChatGPT Enterprise and Edu users will get access next week. Developers can also use the API for prototyping.