OpenAI concluded its 12-day “shipmas” event with a major announcement: the launch of o3, the successor to the o1 reasoning model introduced earlier this year. Like its predecessor, o3 comes as a model family, including a smaller, task-specific version called o3-mini.
OpenAI claims that under certain conditions, o3 approaches artificial general intelligence (AGI), though with notable caveats. Interestingly, the model was named o3 instead of o2, reportedly to avoid trademark conflicts with the British telecom company O2, as hinted by CEO Sam Altman during a recent livestream.
While o3 and o3-mini are not yet widely available, safety researchers can now preview o3-mini. A broader release of o3-mini is expected by late January, with o3 to follow. However, this timeline contrasts with Altman’s earlier suggestion that a federal testing framework should be in place before releasing advanced reasoning models.
Reasoning models like o1 and o3 differ from traditional AI by “fact-checking” their responses, making them more reliable in areas like physics, science, and mathematics. This process takes extra time, ranging from seconds to minutes, but it allows the models to reason through tasks using what OpenAI calls a “private chain of thought.” New to o3 is an adjustable reasoning time feature, which improves performance at higher compute settings.
Benchmarks suggest that o3 is a significant step forward. It scored 87.5% on the ARC-AGI test in high-compute mode, tripling o1’s performance at its lowest setting. The model also excelled in programming and academic benchmarks, setting records in tasks like the SWE-Bench Verified and the American Invitational Mathematics Exam. However, these results come from internal evaluations and await validation from external organizations.
.jpg "OpenAI")
The o3 announcement arrives amid a surge of reasoning model development by competitors like Google, DeepSeek, and Alibaba, reflecting the industry’s shift toward novel methods for improving AI. However, reasoning models face challenges, such as high costs and uncertain long-term progress.
Coinciding with o3’s debut is the departure of Alec Radford, a leading OpenAI scientist and co-author of the paper that introduced the GPT series of generative AI models.