OpenAI has unveiled a new model for its products, arriving for users near the end of January, 2025: It's called o3 (we seem to have jumped over o2), and it promises another significant step forward in AI reasoning. According to its developers, it will make tools like ChatGPT better than ever at programming and working out math problems.
OpenAI CEO Sam Altman described o3 as "incredibly smart" in the video announcing the model, released as part of his company's "12 Days of OpenAI" promotion over the holiday season. The model is undergoing a variety of safety tests before it launches in full—first likely only for paying ChatGPT Plus users.
The o3 model is more than 20 percent better than the previous o1 model at coding, per the SWE-bench Verified benchmark, OpenAI says. It also scores strongly on math and science problems, at least according to benchmark tests—like o1, the o3 model is trained to think and reason before it answers, rigorously testing its responses for accuracy. OpenAI will also release a smaller, faster o3-mini model alongside the main update.
We won't know just how good o3 is until users can actually test it, but we already have an idea of what o3 can do because it's been tested against the well-known Abstraction and Reasoning Corpus (ARC) challenge, designed to track AI's progress towards Artificial General Intelligence (AGI)—the somewhat contentious point at which AI cognitive capabilities pass those of humans.
This challenge gets AI to come up with new approaches to problems, rather than just relying on its memory, and involves a series of visual tasks for models to complete. They must match patterns in colored grids, exercises intended to be easy for people to complete without any training, but hard for AI to figure out.
Within the computing power boundaries of the ARC test, o3 scored 75.7%. That's way above the 5% achieved by the GPT-4o model, currently the best ChatGPT model available to free users. While we're still some way short of AGI (the model is still below human scores, and couldn't complete all the tasks), that's an impressive step up.
"OpenAI's new o3 model represents a significant leap forward in AI's ability to adapt to novel tasks," writes François Chollet, the software engineer who designed the ARC test. "This is not merely incremental improvement, but a genuine breakthrough, marking a qualitative shift in AI capabilities compared to the prior limitations of LLMs."
Predictably, OpenAI didn't talk about the energy demands of AI, the ethics of training AI on publicly available data that may be copyrighted, or the tendency for these models to hallucinate wrong answers—while mistakes should be fewer because of o3's extra thinking time, they won't be eradicated. What the company did mention is an expansion of its safety testing program, designed to prevent these models from being used for malicious purposes.
The ability for AI models to truly "think" or "reason"—or at least attempt some approximation of those human capabilities—will no doubt continue to be discussed as AI development progresses. Google has also just unveiled its Gemini 2.0 model, which brings with it improved reasoning.
