OpenAI is updating the AI mannequin powering Operator, its AI agent that may autonomously browse the net and use sure software program inside a cloud-hosted digital machine to satisfy customers’ requests.
Quickly, Operator will use a mannequin based mostly on o3, one of many newest in OpenAI’s o sequence of “reasoning” fashions. Beforehand, Operator relied on a customized model of GPT-4o.
By many benchmarks, o3 is a much more superior mannequin, significantly on duties involving math and reasoning.
“We’re changing the present GPT‑4o-based mannequin for Operator with a model based mostly on OpenAI o3,” OpenAI wrote in a weblog submit. “The API model [of Operator] will stay based mostly on 4o.”
Operator is one amongst many agentic instruments launched by AI corporations in current months. Corporations are racing to make extremely subtle brokers that may reliably perform chores kind of with out supervision.
Google gives a “computer use” agent via its Gemini API that may equally browse the net and take actions on behalf of customers, in addition to a extra consumer-focused providing referred to as Mariner. Anthropic’s fashions are additionally in a position to perform computer tasks, together with opening recordsdata and navigating net pages.
Based on OpenAI, the brand new Operator mannequin, referred to as o3 Operator, was “fine-tuned with further security knowledge for laptop use,” together with datasets designed to “train the mannequin [OpenAI’s] choice boundaries on confirmations and refusals.”
OpenAI has launched a technical report exhibiting o3 Operator’s efficiency on particular security evaluations. In comparison with the GPT-4o Operator mannequin, o3 Operator is much less prone to refuse to carry out “illicit” actions and seek for delicate private knowledge, and fewer prone to a type of AI assault generally known as immediate injection, per the technical report.
“o3 Operator makes use of the identical multi-layered strategy to security that we used for the 4o model of Operator,” OpenAI wrote in its weblog submit. “Though o3 Operator inherits o3’s coding capabilities, it doesn’t have native entry to a coding surroundings or terminal.”
