Key information
- OpenAI introduced o3, the successor to its initial “reasoning” model.
- The new model family includes both o3 and o3-mini, which are currently open for public safety testing.
- o3 demonstrated significant improvement over its predecessors, achieving a 23 percentage point advantage over o1 on OpenAI's SWE-Bench Verified benchmark.
OpenAI introduced o3, the successor to its initial “reasoning” model, at the conclusion of its “12 Days of OpenAI” product launch event. The new model family includes both o3 and o3-mini. Although not immediately available to the public, these models are currently open for public safety testing.
During a live announcement, OpenAI CEO Sam Altman highlighted that this is the start of a new era of AI where complex reasoning tasks become increasingly feasible. He explained the decision not to use the “o2” designation out of respect for Telefónica, a mobile network operator, and recognizing OpenAI's history with model naming.
New features and performance
For the first time, OpenAI is inviting external security researchers to preview these models. Altman shared that o3-mini will be released towards the end of January, followed shortly after by the full o3 model. Compared to its predecessors, o1 and o1-mini, o3 demonstrated a significant improvement. It achieved a 23 percentage point advantage over o1 in OpenAI's SWE-Bench Verified assessment and achieved a Codeforces score of 2727, even surpassing the score achieved by OpenAI's chief scientist.
Comparison with previous models
OpenAI initially launched the full version of its o1 model during the first day of its “12 Days of OpenAI” promotional campaign. Along with this announcement, it introduced ChatGPT Pro, a new $200 monthly subscription for ChatGPT that includes an advanced version of o1 known as “o1 pro mode.”
If you want access to all articles, subscribe here!
