OpenAI's Operator gets an upgrade with new o3 reasoning model
What's the story
OpenAI has announced an upgrade to the AI model powering its autonomous web agent, Operator.
The new version will be based on o3, a cutting-edge addition to OpenAI's series of "reasoning" models.
This upgrade replaces the previous custom version of GPT-4o that Operator used, according to a blog post by OpenAI.
Model performance
o3 model excels in math and reasoning tasks
The decision to upgrade to the o3 model stems from its better performance in a number of benchmarks, especially those involving math and reasoning tasks.
OpenAI said in its blog post that while the API version of Operator will continue to use the 4o model, they are replacing the current GPT-4o-based model for Operator with an OpenAI o3-based version.
AI competition
Operator joins a growing list of advanced AI agents
The upgraded Operator joins a growing list of advanced AI agents from other tech companies. These agents are capable of performing tasks with little to no supervision.
Google, for instance, has launched a "computer use" agent via its Gemini API, capable of browsing the web and taking actions on the user's behalf.
Anthropic also has models capable of performing computer tasks such as opening files and navigating web pages.
Twitter Post
Operator is now more persistent and accurate
Now Operator is more persistent and accurate when interacting with the browser, improving overall task success rate. Responses are also more clear, thorough and better structured. More detailed evaluations
— OpenAI (@OpenAI) May 23, 2025can be found in the release notes: https://t.co/3gSEvoOizOpic.twitter.com/HLA7VZM21U
Safety measures
New Operator model enhanced with safety data
The latest version of the Operator model, o3 Operator, has been fine-tuned with more safety data for computer use.
This includes data sets designed to teach the model, OpenAI's decision boundaries on confirmations and refusals.
A technical report released by OpenAI shows that compared to the GPT-4o Operator model, o3 Operator is less likely to perform illicit activities or search for sensitive personal data.
It is also less vulnerable to an AI attack known as prompt injection.
Information
o3 Operator inherits coding capabilities but lacks native access
OpenAI has also clarified in its blog post that while the o3 Operator inherits o3's coding capabilities, it does not have a native access to a coding environment or terminal.