OpenAI upgrades Operator with o3 model for enhanced reasoning, safety

OpenAI is updating the artificial intelligence (AI) model powering Operator, its AI agent that can autonomously browse the web and interact with certain software inside a cloud-hosted virtual machine to carry out user requests.

Operator will soon run on a model based on o3, one of the latest in OpenAI’s o series of “reasoning” models. Previously, Operator relied on a customised version of GPT-4o.

By several benchmarks, o3 is a more advanced model, particularly on tasks requiring mathematical ability and reasoning.


“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 based on 4o.”

Operator is part of a growing set of agentic tools developed by AI firms as they compete to build agents capable of reliably performing digital tasks with minimal supervision.

Google offers a similar agent through its Gemini API, which can browse the web and take actions on users’ behalf. It also offers a consumer-facing version called Mariner. Anthropic’s models can perform various computer tasks as well, including opening files and navigating webpages.

According to OpenAI, the upgraded Operator model, dubbed o3 Operator, was “fine-tuned with additional safety data for computer use,” using datasets designed to “teach the model (OpenAI’s) decision boundaries on confirmations and refusals.”

The company has released a technical report detailing o3 Operator’s performance in safety evaluations. Compared to the GPT-4o version, the new model is less likely to carry out illicit activities, search for sensitive personal data or fall prey to prompt injection, a common AI attack technique.

“o3 Operator uses the same multi-layered approach to safety that we used for the 4o version of Operator,” OpenAI wrote in its blog post. “Although o3 Operator inherits o3’s coding capabilities, it does not have native access to a coding environment or terminal.”