Operator, an AI agent from OpenAI that can browse the web on its own and use certain applications in a virtual machine stored in the cloud to respond to user requests, is getting an update to its AI model.
One of the newest in OpenAI’s o family of “reasoning” models, o3, will soon be used by Operator. Operator previously used a customized version of GPT-4o.
O3 is a significantly more sophisticated model by many standards, especially when it comes to math and reasoning problems.
In a blog post, OpenAI stated, “We are replacing the existing GPT‑4o-based model for Operator with a version based on OpenAI o3.” “Operator’s API version will continue to be based on 4o.”
In recent months, AI companies have produced a number of agentic tools, including Operator. Businesses are rushing to create extremely complex agents that can consistently complete tasks largely unsupervised.
Through its Gemini API, Google provides a “computer use” agent that can similarly explore the web and do actions on behalf of users. Mariner is a more consumer-focused product. Additionally, Anthropic’s models can navigate websites and open files, among other computer functions.
In order to “teach the model [OpenAI’s] decision boundaries on confirmations and refusals,” the new Operator model, known as o3 Operator, was “fine-tuned with additional safety data for computer use,” according to OpenAI.
A technical report from OpenAI details how well o3 Operator performed on particular safety assessments. According to the technical paper, o3 Operator is less likely than the GPT-4o Operator model to decline to engage in “illicit” activities and look for private information. It is also less vulnerable to prompt injection, a type of AI attack.
In a blog post, OpenAI stated that “o3 Operator employs the same multi-layered approach to safety that we used for the 4o version of Operator.” “O3 Operator does not have native access to a coding environment or terminal, even though it inherits o3’s coding capabilities.”