On the final day of its 12-day “shipmas” event, OpenAI announced its newest AI model: O3. O3 is the successor to the o1 reasoning model launched earlier this year and promises a major leap in AI capabilities. OpenAI also introduced o3-mini designed for specific, high-efficiency tasks.
You might be wondering why the successor to O1 isn’t called O2. The company skipped making it O2 to avoid trademark conflicts with the British telecom giant O2. Sam Altman, OpenAI CEO, confirmed the decision during a livestream.
In tests like the American Invitational Mathematics Examination, o3 demonstrated an accuracy of 96.7%. It outperformed o1’s 83.3% accuracy. Notably, o3 also performed exceptionally well on ARC-AGI, a benchmark designed to evaluate an AI’s ability. On a low-compute setting, o3 scored 75.7%, and with enhanced processing power, it achieved an impressive 87.5%.
Altman also claims that o3 model is incredible at coding and benchmarks confirm it. o3 surpasses o1 by 22.8 percentage points on SWE-Bench Verified and achieves a Codeforces rating of 2727.
The model also sets a records on challenging tests like EpochAI’s Frontier Math. It solved 25.2% of problems where no other model exceeds 2%.
The o3-mini model incorporates OpenAI’s Adaptive Thinking Time API. This allows users to select reasoning modes—Low, Medium, or High—depending on the complexity of the problem. o3-mini achieves performance on par with the o1 model and claims to be a cost-effective alternative for many applications.
Altman also emphasized OpenAI’s commitment to safety. He confirmed that thorough testing and risk mitigation frameworks are crucial before deploying reasoning models to the public.
O3 and O3 mini models will be released next year. Applications for early access are now open until January 10, 2025, on the OpenAI website. If you are interested, you must fill out an online form. Selected users will get access to the o3 and O3-mini models.
