OpenAI Switches Operator to New o3 Model, Keeps GPT-4o for API

OpenAI launched Operator, a browser-based assistant powered by its Computer Using Agent (CUA) model. The model performs tasks online by clicking, typing, and scrolling—much like a human user.

Now, OpenAI has swapped out the original GPT?4o model behind Operator with a version based on its newer o3 system. The API version will continue using GPT?4o, so developers won’t see changes there.

The o3-based Operator keeps the same safety layers as before, originally detailed in OpenAI’s first Operator System Card. Engineers also fine-tuned this new version with extra safety training, focused specifically on computer interaction. That training helps the model recognize when to confirm or refuse actions, reducing the chances of unexpected behavior.

Also read: Open AI’s Sora is just too good that it might’ve been trained in Unreal 5

While the model brings o3’s coding knowledge, it still can’t access a live coding environment or a terminal. It stays strictly within the browser space.

OpenAI says this shift continues its research into AI agents that can interact with websites reliably and safely. Operator remains a research preview for now, giving OpenAI more space to study and refine agent behavior in real-world web environments without full public rollout.