OpenAI o3 and o4-mini Release Notes
OpenAI o3 and OpenAI o4-mini are reasoning models that integrate state-of-the-art reasoning with comprehensive tool capabilities, including web browsing, Python, image and file analysis, image generation, canvas, automations, file search, and memory. These models are designed to solve complex mathematical, coding, and scientific challenges while providing strong visual perception and analysis.
Tool Integration in Chain-of-Thought Reasoning
OpenAI o3 and o4-mini augment their reasoning process by utilizing tools directly within their chains of thought. This allows the models to perform actions such as cropping or transforming images, searching the web, or executing Python code to analyze data during their internal deliberation process before arriving at a final answer.
Training and Safety Alignment
The o-series models are trained using large-scale reinforcement learning on chains of thought. This training approach enables "deliberative alignment," where the models can reason about safety policies in context when responding to potentially unsafe prompts, thereby improving overall model robustness and safety.
Preparedness Framework and Risk Evaluation
This release is the first to be evaluated under Version 2 of OpenAI's Preparedness Framework. The Safety Advisory Group (SAG) reviewed the evaluations across three Tracked Categories:
- Biological and Chemical Capability: The models did not reach the High threshold for risk.
- Cybersecurity: The models did not reach the High threshold for risk.
- Cybersecurity: The models did not reach the High threshold for risk AI Self-improvement.
OpenAI reports that o3 and o3-mini are compliant with the safety thresholds established by the Safety Advisory Group.