OpenAI has announced the testing of new artificial intelligence models named o3 and o3 mini, which are capable of tackling more complex tasks. A review of the new models is available on the company's YouTube channel.
OpenAI's CEO Sam Altman described the o3 and o3 mini models as "very, very smart" and "incredibly intelligent." In real programming tasks (SWE Bench), the o3 model achieved an accuracy of 71.7%, which is 20 percentage points higher than the o1 model.
In coding competitions (Codeforces), the model received an ELO rating of 2727, significantly surpassing the previous o1 model (1891). For comparison, OpenAI programmers typically score around 2500, while their top specialists score slightly higher.
In the mathematical test EpochAI Frontier Math, o3 achieved an accuracy of 25.2%, which is more than 12 times better than previous models.
https://twitter.com/__nmca__/status/1870170098989674833
The o3-mini model supports three levels of reasoning: low, medium, and high. The more time the model spends processing a task, the more accurate the result. In low latency mode, o3-mini demonstrates speed on par with GPT-4o, allowing for quick responses.
The new o3 mini models are expected to be released by the end of January 2025, with the full version of o3 following after that. With these models, OpenAI aims to surpass existing language models and attract new users and investments.
OpenAI has already opened the application process for external security researchers to test the new models before their public release. The application period will close on January 10, 2025. Applications for model testing can be submitted via the link.
Earlier, Spot reported that OpenAI launched a tool for generating videos based on text prompts.