The new models, known as o1 and o1-mini, are capable of reasoning about complex tasks and can solve more difficult problems in science, coding, and mathematics than previous models.

OpenAI said the o1 model will be integrated into ChatGPT and its application programming interface (API) from September 12.

According to the company, the o1 model achieved an accuracy rate of up to 83% when solving problems in the qualifying round of the International Mathematical Olympiad, a significant improvement compared to the 13% of the previous GPT-4o model.

Going beyond mathematics, the o1 model has also improved performance in solving competitive programming questions and even surpassed the accuracy of PhD-level scientists on a science test.

One of the key factors that helped the o1 model achieve this success was the application of the “chain-of-thought” technique. This is a method that allows the model to break down complex problems into smaller logical steps.

According to the researchers, using this technique helps improve the performance of AI models when faced with difficult problems.

Notably, OpenAI has automated this process, allowing AI models like o1 to break down and analyze problems without any prompting from the user.

“We trained these models to spend more time thinking before responding, similar to how humans would. Through training, the models learn to refine their reasoning process, try different strategies, and recognize their mistakes,” OpenAI said.