(Credit: OpenAI)
OpenAI released a new AI model, dubbed GPT-o1, the culmination of its rumored project Strawberry.
The model could eventually be available on the free version of ChatGPT, but for now it's only rolling out to paying ChatGPT Plus and Team users, and in the developer API. Developers will also get access to the more cost-efficient o1-mini variant.
How does GPT-o1 differ from its predecessor, GPT-4o, which debuted in May? It takes longer to answer to your questions, which OpenAI argues is a mark of greater intelligence.
"We trained these models to spend more time thinking through problems before they respond, much like a person would," OpenAI says. "Through training, they learn to refine their thinking process, try different strategies, and recognize their mistakes."
It can explore a train of thought, and therefore processes information differently. While GPT-4o notoriously thinks the word "strawberry" only has two "r"s, GPT-o1 can identify that there are three in a demo video, hence the project code name.
GPT-o1 is geared toward researchers, scientists, and developers who work on complex coding and math problems. In internal testing, GPT-o1 solved 83% of questions on the qualifying exam for the International Mathematics Olympiad (IMO), while GPT-4o only solved 13%. The new model also "performs similarly" to PhD students on physics, chemistry, and biology tasks, OpenAI says.
CEO Sam Altman calls GPT-o1 "the beginning of a new paradigm," but admits it's "still flawed, still limited, and still seems more impressive on first use than it does after you spend more time with it." For example, it cannot browse the web or read uploaded files and images. For those everyday tasks, GPT-4o "will be more capable in the near term," OpenAI says.
Though the long-awaited GPT-5 is still under development, OpenAI calls GPT-o1 a critical breakthrough, and possibly a step toward artificial general intelligence (AGI).
"I have always believed that you don't need a GPT-6 quality base model to achieve human-level reasoning performance, and that reinforcement learning was the missing ingredient on the path to AGI," says OpenAI researcher Max Schwarzer.