OpenAI Announces New AI Model, Codenamed Strawberry, That Can Solve Difficult Problems Step by Step
OpenAI made the final major breakthrough in artificial intelligence by increasing the size of its models to dizzying levels, as it Introducing GPT-4 last year. Today, the company announced a new advance that signals a shift in approach—a model that can “reason” rationally across a wide range of difficult problems and is significantly smarter than current AI without needing to scale at a massive scale.
The new model, called OpenAI o1, can solve problems faced by existing AI models, including OpenAI’s most powerful model currently available, GPT-4o. Instead of coming up with an answer in one step like a large language model would, it reasoned through the problem, thinking out loud like a normal person, before coming up with the correct result.
“This is what we consider a new model within these models,” Mira Murati“It’s much better at solving very complex reasoning tasks,” OpenAI’s chief technology officer told WIRED.
The new model is codenamed Strawberry within OpenAI, and the company says it’s not a successor to GPT-4o but rather a complement to it.
Murati says OpenAI is currently building its next major model, GPT-5, which will be significantly larger than its predecessor. But while the company still believes that scale will help unlock new possibilities from AI, GPT-5 will likely also include the reasoning technology introduced today. “There are two models,” Murati says. “The scaling model and this new model. We hope to combine them.”
LLMs typically generate answers from massive neural networks fed large amounts of training data. They can demonstrate remarkable linguistic and logical abilities, but have traditionally struggled with surprisingly simple problems like basic mathematical reasoning.
Murati said the OpenAI o1 uses reinforcement learning, which involves giving the model positive feedback when it gets an answer right and negative feedback when it doesn’t, to improve its reasoning process. “The model sharpens its thinking and refines the strategies it uses to come up with answers,” she said. Reinforcement learning has allowed computers to play games with super skills and do useful work like computer chip design. This technique is also an important factor in making LLM a useful and well-functioning chatbot.
Mark Chen, vice president of research at OpenAI, demonstrated the new model to WIRED, using it to solve several problems that the company’s previous model, GPT-4o, couldn’t solve. These included an advanced chemistry question and the following tricky math puzzle: “A princess will be the same age as the prince when the princess is twice as old as the prince when the princess’s age is half the sum of their current ages. What are the ages of the prince and the princess?” (The correct answer is that the prince is 30 and the princess is 40).
“The [new] “This model is learning to think for itself, rather than trying to mimic how humans think,” as a typical LLM would, Chen said.
OpenAI said its new model performed significantly better on a number of problem sets, including those focused on coding, mathematics, physics, biology, and chemistry. On the American Invitational Mathematics Examination (AIME), a test for math students, GPT-4o solved an average of 12 percent of problems, while o1 got 83 percent correct, the company said.
