OpenAI is updating the AI model that powers the Operator agent

OpenAI is updating the artificial intelligence model that powers Operator, an AI agent capable of autonomously browsing web pages and using specific software in a cloud virtual machine to carry out user requests.

Soon, Operator will use a model based on o3 — one of the newest models in OpenAI’s series of “reasoning” models. Previously, Operator ran on a specialized version of GPT‑4o.

By many metrics, o3 is a significantly more advanced model, particularly in tasks involving mathematics and logical reasoning.

Operator is one of many agent tools released by AI companies in recent months. Currently, there is a race to create highly sophisticated agents capable of performing tasks with minimal human intervention.

Google offers a computer use agent through its Gemini API — it can also browse web pages and perform actions on behalf of the user. Additionally, Google has a more consumer-oriented product called Mariner. Anthropic’s models can also perform computer tasks, including opening files and navigating web pages.

According to OpenAI, the new Operator model, named o3 Operator, was fine-tuned using additional data on “computer use safety,” including datasets designed to “train the model on the boundaries of accepting and refusing to perform actions.”

OpenAI released a technical report on the performance of o3 Operator on certain safety checks. Compared to the GPT‑4o Operator model, the o3 Operator model is less likely to agree to perform “illegal” actions, search for personal data, and is less susceptible to attacks via so-called prompt injection.

o3 Operator uses the same multi-layered approach to security as the 4o-based version. Although o3 Operator inherited the software capabilities of the o3 model, it does not have native access to a programming environment or terminal.

Source: TechCrunch