OpenAI’s Operator agent is getting a significant model change inside ChatGPT. The computer-using agent now runs on a new model based on the o3 architecture, a move aimed at making web automation more precise, more structured and more reliable.
The upgraded Operator is available worldwide in ChatGPT-Pro as a research preview. API usage, however, is still based on GPT-4o.
What changed for Operator
Operator is OpenAI’s Computer-Using Agent, or CUA. Its role is to interact with websites in ways that resemble a human using a browser: scrolling, clicking and typing text to complete tasks online.
The new o3-based model replaces the previous GPT-4o-based version of Operator in ChatGPT. OpenAI first introduced Operator as a research preview in January 2025, with the broader goal of building an AI agent able to perform web-based actions in a human-like way.
That matters because browser-based work is often made of many small steps. A user may need an agent to read a page, move through forms, identify the next action and continue without losing the thread of the task. The source describes the upgrade as an effort to make Operator more precise, more structured and more successful on the web.
More precise browser control
OpenAI says the switch to o3 is designed to make Operator more robust and effective at completing web tasks. The model is described as interacting more precisely with browsers, which is central to the product’s purpose.
In practical terms, browser control is not just about generating a good answer. The agent must also understand the visible state of a website and choose actions that move a task forward. The source specifically points to scrolling, clicking and typing text as the kinds of actions Operator can perform.
OpenAI also says the upgraded model produces responses that are better structured and more comprehensive. That is important for an agent whose work may involve both taking actions and explaining what happened. Structure can help users understand the outcome of a workflow, especially when a task has multiple steps.
Internal testing shows Operator now succeeds more often at complex workflows. OpenAI also says the model sets the standard in benchmarks like OSWorld and WebArena, while user tests show better response quality than its predecessor.
Why o3 is being used here
The o3 Operator model is built on the same architecture as other o3 models. But it has been specifically trained to operate computer interfaces, according to OpenAI.
That distinction is important. A model that can reason well still needs training for the environment where it will act. In Operator’s case, that environment is the browser and the websites inside it.
The source also notes a boundary: although o3 Operator inherits o3’s coding capabilities, OpenAI says it does not have direct access to coding environments or terminals. The upgrade is therefore focused on browser use, not on giving the agent direct access to development tools.
For ChatGPT-Pro users, the current result is a research preview of an o3-based Operator inside ChatGPT. For developers using the API, the source says usage is still based on GPT-4o.
Safety is part of the upgrade
OpenAI says the new model was fine-tuned with additional security data. The goal is to help it learn when it should provide confirmations or refusals.
This matters because browser automation creates a specific kind of risk. These agents must analyze website content and treat parts of that content as instructions. In effect, a web page can become something the agent interprets while deciding what to do next.
The source highlights the danger of malicious websites designed to manipulate an agent. For example, an attacker could create a site that tries to trick the agent into unwanted actions, such as entering sensitive information into fake login forms.
That makes safety behavior part of the core product, not a side feature. A browser agent needs to know how to proceed when a page asks for an action, but it also needs to recognize moments when it should pause, ask for confirmation or refuse.
What this signals for web agents
The Operator upgrade shows how OpenAI is positioning browser automation as a specialized task. The agent is not only being judged by whether it can write a response, but by whether it can complete workflows on websites with greater precision.
The move from a GPT-4o-based Operator to an o3-based model also separates the ChatGPT research preview from API usage, which remains on GPT-4o according to the source. That creates a clear distinction between the current ChatGPT-Pro experience and what is available through the API.
For users watching the development of AI agents, the key point is straightforward: OpenAI is trying to make Operator better at the browser itself. The company says the new model improves precision, structure, response quality and success on complex workflows, while adding security-focused training for confirmations and refusals.
Operator remains a research preview, but the upgrade points to the direction of the product: agents that can handle web tasks with more reliable control and more careful decisions about when to act.