On the final day of the 12-day-long "Shipmas" event, OpenAI introduced its latest reasoning models, named o3 and o3-mini. These models were unveiled as a preview, with no immediate plans for public release. OpenAI CEO Sam Altman announced that the o3-mini model is expected to launch by the end of January, with the full version of o3 to follow at a later date. Altman believes these advanced large language models (LLMs) have the potential to outperform existing ones, attracting both new users and investors.
In preparation for the launch, OpenAI has invited researchers to apply for early testing of these models before they are publicly available. The application window for external researchers will remain open until January 10. Notably, OpenAI had previously released o1, initially known as Project Strawberry, in September of this year. Interestingly, the company appears to be skipping the o2 designation entirely, reportedly to avoid a potential copyright conflict with the UK-based telecommunications company O2.
OpenAI o3 Outperforms Its Predecessor
The o3 model has reportedly set new benchmarks, outperforming its predecessors in multiple domains. It achieved a 22.8 per cent improvement in SWE-Bench verified coding tests and even outperformed OpenAI’s Chief Scientist in competitive programming challenges.
Additionally, the model excelled in AIME 2024, one of the most challenging math competitions, scoring an impressive 87.7 per cent on expert-level science problems. When tackling the most complex math and reasoning challenges, o3 solved 25.02 per cent of the problems—far surpassing other models, which struggled to exceed a 2 per cent success rate.
While the o1 models demonstrated strong reasoning capabilities and handled intricate problems effectively, the upcoming o3 and o3-mini models promise to deliver even greater performance. At present, these new models are undergoing rigorous internal safety testing.