OpenAI has announced the development of its next generation of reasoning models, the o3 family. These models, which include o3 and o3 Mini, demonstrate significant advancements in AI capabilities, particularly in problem-solving and adaptability.
Benchmark tests reveal that the o3 models outperform existing models by a substantial margin. The o3 model achieved a remarkable 75.7% on the ARC-AGI Semi-Private Evaluation, with a high-compute configuration reaching 87.5%. On other benchmarks, such as EpochAI Frontier Math and Codeforces, the o3 models also demonstrated superior performance.
The o3 Mini model will offer users three reasoning levels: High, Medium, and Low. This allows for flexibility in balancing speed and accuracy, catering to different user needs.
OpenAI is currently conducting safety and security testing for the o3 models and has invited researchers to participate in the process. The o3 models are expected to be publicly available in 2025.
