Openai is updating its AI model power operator. It updates AI agents that can autonomously browse the web and satisfy user requests using specific software within cloud-hosted virtual machines.
Soon, the operator will use the latest O3-based model of OpenAI’s O-series “inference” model. Previously, operators relied on custom versions of GPT-4O.
With many benchmarks, O3 is a much more advanced model, especially in tasks that involve mathematics and inference.
“We are replacing our existing GPT‑4o-based models for operators with versions based on Openai O3,” Openai wrote in a blog post. “The (operator’s) API version is based on 4o.”
Operators are one of many agent tools that AI companies have released in recent months. Companies compete to create very sophisticated agents who can ensure that they perform more or less chores without oversight.
Google provides a “computer use” agent via the Gemini API. It also browses the web, performs actions on your behalf, and offers a more consumer-centric offering called Mariner. Anthropic models can also perform computer tasks, such as opening files and navigation of web pages.
According to Openai, the new operator model, called the O3 operator, is “finely tuned with additional safety data for computer use,” including a dataset designed to “tell the model (OpenAI) decision boundaries regarding confirmation and rejection.”
Openai has released a technical report showing the performance of O3 operators on specific safety ratings. Compared to the GPT-4O operator model, O3 operators are less likely to reject “illegal” activities or search for sensitive personal data, and are less susceptible to AI attacks known as rapid injection, according to technical reports.
“O3 operators safely use the same multi-layer approach they used for the 4O version of the operator,” Openai wrote in a blog post. “O3 operators inherit the coding functionality of O3, but do not have native access to the coding environment or terminal.”