OpenAI is giving an upgrade to Operator, its AI agent that makes life easier by browsing the web or using specific software on cloud-based virtual machines based on user instructions.

Operator will soon have a brain upgrade with the o3 reasoning model, a newer member of OpenAI’s 'o' family of reasoning models. Previously, Operator was powered by GPT-4o, but the o3 model promises superior performance, particularly in tasks involving math and reasoning.

In an OpenAI blog post, the team announced that they're moving to an o3 based model for Operator, while the API version continues to use 4o.

Operator isn't alone; it's part of the recent wave of AI tools known as agent tools that aim to perform tasks with minimal supervision. Google offers a parallel tool via its Gemini API, which, like Operator, can browse the web and act on behalf of users, and also has a more consumer-focused offering named Mariner. Anthropic's models are similarly skilled, capable of specific computer tasks like opening files and navigating web pages.

OpenAI's new o3 Operator has been refined with additional safety data for computer use, resulting in enhanced decision-making parameters.

OpenAI has shared a report detailing how the o3 Operator fared on specific safety evaluations. According to this report, the o3 model is at a lower risk of refusing to perform “illicit” activities and accessing sensitive personal data. It's also more resilient to a commonly used AI attack, prompt injection.

OpenAI, in its blog post, states that for safety measures, o3 Operator follows the same safety approach as the 4o version of Operator. However, while o3 Operator inherits o3’s coding prowess, it doesn't directly interact with a programming environment or terminal.