Addendum to OpenAI o3 and o4-mini system card: OpenAI o3 Operator

# Addendum to OpenAI o3 and o4-mini system card: OpenAI o3 Operator Source: [https://openai.com/index/o3-o4-mini-system-card-addendum-operator-o3/](https://openai.com/index/o3-o4-mini-system-card-addendum-operator-o3/) In January 2025, we shipped Operator, a product to serve our Computer Using Agent \(CUA\) model as a research preview\. CUA is an agentic model that can use the web to perform tasks for the user\. Using its own browser, it can look at a webpage, and interact with it much like a human would by typing, clicking, scrolling and more\. We are replacing the existing GPT‑4o‑based model for Operator with a version based on OpenAI o3\. The API version will remain based on 4o\. o3 Operator uses the same multi\-layered approach to safety that we used for the 4o version of Operator and described in our original[Operator System Card⁠\(opens in a new window\)](https://cdn.openai.com/operator_system_card.pdf)\. Compared with other models in the o3 family, o3 Operator was fine\-tuned with additional safety data for computer use, including safety datasets designed to teach the model our decision boundaries on confirmations and refusals\. Although o3 Operator inherits o3’s coding capabilities, it does not have native access to a coding environment or Terminal\.