OpenAI is updating the artificial intelligence model that powers Operator, an AI agent that can autonomously browse the web and use specific software on a virtual machine in the cloud to fulfill user requests.
Soon, Operator will be using a model based on o3, one of the latest o-series “reasoning” models from OpenAI. Previously, it relied on a custom version of GPT-4o.
In many ways, o3 is a much more advanced model, especially in math and reasoning tasks.
“We are replacing the existing GPT-4o-based model for Operator with a version based on OpenAI o3,” OpenAI wrote in a blog post. “The API version [of Operator] will remain 4o-based.”
Operator is one of many agent tools released by AI companies in recent months. Companies are competing to create high-tech agents that can reliably do their homework more or less unsupervised.
Google offers a “computer use” agent through its Gemini API, which can also browse the web and perform actions on behalf of users, as well as a more consumer-oriented offering called Mariner. Anthropic models are also capable of performing computer tasks, including opening files and navigating web pages.
According to OpenAI, the new operator model, dubbed the o3 Operator, has been “refined with additional computer security data,” including datasets designed to “train the [OpenAI] model on confirmation and denial decision boundaries.”
OpenAI has released a technical report demonstrating the performance of o3 Operator in specific security assessments. Compared to the GPT-4o Operator model, o3 Operator is less likely to refuse to perform “illegal” actions and search for sensitive personal data, and is less susceptible to a form of AI attack known as “quick input,” according to the technical report.
“o3 Operator uses the same layered approach to security that we used for version 4o Operator,” OpenAI wrote in a blog post. “While o3 Operator inherits the coding capabilities of o3, it does not have native access to the coding environment or terminal.”
