OpenAI is rolling out a major upgrade to its autonomous software agent, Operator, transitioning the underlying model from GPT-4o to a version based on its latest o3 architecture, the company said in a blog post.
Operator, which can browse the web and interact with software inside a cloud-hosted virtual machine on behalf of users, is now powered by “o3 Operator,” a variant of the o3 model series that OpenAI describes as more advanced in tasks involving mathematical reasoning and complex decision-making.
“We are replacing the existing GPT‑4o-based model for Operator with a version based on OpenAI o3,” the company wrote, noting that the API version of Operator will continue to use GPT-4o.
The o3 series represents OpenAI’s newest generation of “reasoning” models. According to a technical report published alongside the announcement, o3 Operator demonstrates stronger resistance to prompt injection attacks and reduced likelihood of executing potentially harmful or privacy-invasive tasks when compared to the previous model.
The upgraded model was fine-tuned with additional safety data specific to computer use, including guidance on when to request user confirmations or issue refusals. “o3 Operator uses the same multi-layered approach to safety that we used for the 4o version of Operator,” OpenAI wrote.
Although o3 Operator leverages o3’s enhanced coding capabilities, it does not have native access to a terminal or local coding environment, the company clarified.
OpenAI’s move reflects a broader industry trend toward building sophisticated agentic systems capable of performing digital tasks autonomously. Google’s Gemini API includes a “computer use” agent, while Anthropic’s Claude models can execute file and browser-based actions.
These offerings mark a competitive push among AI developers to create highly capable and safe agents that can function with minimal user oversight, a key step toward broader adoption in enterprise and productivity applications.
