OpenAI has released its o3 and o4-mini models, which combine state-of-the-art reasoning capabilities with full tool usage including web browsing, Python execution, and image analysis. These models are trained using large-scale reinforcement learning on chains of thought to enhance their ability to solve complex math, coding, and scientific challenges.
- The models utilize tools within their thought process to augment capabilities, such as cropping images or analyzing data via Python.
- OpenAI's Safety Advisory Group reviewed the models under Version 2 of its Preparedness Framework.
- Evaluations determined that neither model reaches the High threshold in Biological and Chemical Capability, Cybersecurity, or AI Self-improvement.
The release marks the first launch under the updated framework, with the company noting that advanced reasoning provides new avenues for improving safety and robustness through deliberative alignment.