OpenAI has announced its new simulated reasoning models. They are named o3 and o3-mini and build on the earlier o1 models.
OpenAI’s new o3 and o3 models have been announced, but are not yet becoming generally available. First, they will be made publicly available for public security testing and research.
Simulated reasoning
The models build on the o1 models launched earlier this year. They use a “private chain of thought. The model pauses briefly to examine internal dialogue and plan ahead for responding. That form is called “simulated reasoning,” surpassing the thinking power of simple LLMs. Google is also launching its Gemini 2.0 Flash Thinking model that uses that same method of simulated thinking.
Remarkably, the o3 model achieved high scores on benchmarks such as ARC-AGI and the American Invitational Mathematics Exam. The model scored 87.5% in high-compute tests, comparable to human performance.
According to CEO Sam Altman, the launch of o3-mini is scheduled for late January, followed by the fully rolled-out o3 model shortly thereafter. The o3-mini model includes an adaptive thinking feature that offers low, medium and high processing speeds.
Also read: OpenAI’s GPT-5 model ‘seems unable to meet expectations’