OpenAI introduces its new o3 reasoning AI models to a limited number of users

OpenAI has unveiled the o3 family of AI models, which succeed the previous o1 "reasoning" model. The new series includes the o3 and a smaller version, o3-mini, designed for specific tasks.

The o3 models show significant performance enhancements, including a 22.8% improvement in coding tests on SWE-Bench compared to o1, outperforming OpenAI’s Chief Scientist in coding challenges, and achieving near-perfect scores in the AIME 2024 math competition. They also scored 87.7% on GPQA Diamond expert-level science problems and solved 25.2% of complex reasoning challenges, a substantial increase from earlier models' sub-2% scores.

OpenAI has also introduced a new safety paradigm called deliberative alignment, which involves models processing safety-related decisions step-by-step to improve adherence to safety policies. While not yet publicly available, OpenAI is inviting researchers to begin testing the o3-mini immediately, with a preview of the o3 model to be released later. The company plans to launch o3-mini by the end of January, followed by the release of the main o3 model.