OpenAI Updates Operator With o3 Model for Safer Web-Based Agent Tasks

OpenAI has updated Operator—a research preview released in January 2025 that lets users assign tasks to an AI agent capable of using the web like a person. This agent, called a Computer Using Agent (CUA), can browse websites, read content, click buttons, fill out forms, and scroll through pages much like a human user.

Until now, Operator was powered by a model based on GPT‑4o. OpenAI is now replacing that version with a new variant based on its o3 model family, bringing improved safety features tailored to web-based interaction. (The API version of Operator will continue to run on GPT‑4o.)

What Is Operator?

In this version, Operator is a browser-using AI agent that can carry out tasks on the web—reading pages, clicking buttons, filling out forms, and navigating sites on the user's behalf. It uses a built-in browser and takes actions in real time, acting much like a person would.

While this update focuses on web interaction, Operator is part of a broader research effort exploring how AI agents can take actions across digital systems—not just respond to prompts, but complete real-world tasks using tools, interfaces, and structured decision-making.

This hands-on interaction makes Operator a testbed for “agentic” behavior—where AI systems move from passive assistance to active task completion.

What’s New in the o3 Version

The most important change in this update is the addition of new safety tuning designed specifically for web-based tasks. OpenAI trained the o3-based Operator on datasets that reinforce clear boundaries around what actions it should confirm, refuse, or avoid when using the web on a user's behalf.

This version keeps the same core functionality—browsing, clicking, scrolling, and typing—but introduces more structured guardrails for how it behaves during those tasks.

While the model itself has changed from GPT‑4o to a member of the o3 family, that shift matters less for raw performance and more for how the agent is shaped to act safely and predictably in complex environments like the open internet.

And while it inherits strong coding abilities from o3, this version of Operator does not include direct access to a terminal or coding environment.

You can view the updated safety details in the o3 Operator System Card addendum (PDF).

What This Means

This update reflects a deeper shift in how OpenAI is developing agent-style systems—not just to carry out tasks, but to do so within real-world constraints. By running Operator on an o3-based model fine-tuned specifically for web interaction, OpenAI is testing how an AI agent can behave safely and predictably in open-ended environments like the internet.

That matters because browsing the web is inherently messy. Websites change, information varies in quality, and the actions required to complete a task aren’t always obvious. Operator’s ability to interact with pages like a human—clicking, typing, navigating—offers a new way to automate tasks that previously required direct human involvement. But it also raises new challenges around safety, trust, and oversight.

The o3 version of Operator represents a step toward meeting those challenges. With added training focused on when to act and when to pause, the agent becomes more capable of making sound decisions—even when instructions aren’t perfectly clear.

More broadly, this signals OpenAI’s continued push toward building AI agents that don’t just generate text, but take action—and do so in ways that align with human expectations. As AI tools begin to carry out more complex tasks across digital systems, shaping their behavior through careful tuning and safety data will be just as important as expanding what they can do.

In a future where AI agents will increasingly act on our behalf, how they decide to act may matter as much as what they’re capable of doing.

Editor’s Note: This article was created by Alicia Shapiro, CMO of AiNews.com, with writing, image, and idea-generation support from ChatGPT, an AI assistant. However, the final perspective and editorial choices are solely Alicia Shapiro’s. Special thanks to ChatGPT for assistance with research and editorial support in crafting this article.