OpenAI o3 Operator Update

OpenAI has replaced the GPT-4o-based model powering the Operator research preview with a version based on OpenAI o3. This update transitions the Operator product—a Computer Using Agent (CUA) that can interact with the web via a browser to type, click, and scroll—to a more advanced model architecture while maintaining the same API version based on 4o.

Model Transition and Core Capabilities

OpenAI o3 Operator is a Computer Using Agent (CUA) designed to perform tasks for users by interacting with the web. The model uses its own browser to view webpages and perform human-like interactions, such as clicking, scrolling, and typing. While o3 Operator inherits the general coding capabilities of the o3 family, it does not have native access to a Terminal or a coding environment.

Safety Framework and Fine-Tuning

OpenAI employs a multi-layered safety approach for o3 Operator, consistent with the safety measures described in the original Operator System Card. To ensure the model adheres to decision boundaries regarding confirmations and refusals, o3 Operator was fine-tuned with additional safety data specifically tailored for computer use.

Technical Implementation Details

Despite the model transition to o3, the API version for Operator remains based on 4o. The model is as a research preview available since January 2025.