OpenAI upgrades the AI model powering its Operator agent

OpenAI contributes the AI ​​model of the AI ​​operator, the AI ​​agent who can browse autonomously and use certain software in a cloud-and-hosted virtual machine to meet the requests of users.

Soon the operator will use a model based on O3, one of the latest OpenAi’s O series “reasoning models”. The operator previously relied on a modified version of GPT-4O.

Due to many benchmarks, O3 is a much more advanced model, especially with tasks with mathematics and reasoning.

“We replace the existing GPT-4O-based model for operator with a version based on OpenAI O3,” OpenAi written In a blog post. “The API version [of Operator] remains based on 4O. “

Operator is one of the many agentic tools released by AI companies in recent months. Companies racing to make very advanced agents who can perform reliable jobs without supervision.

Google offers a “computer use” agent via his Gemini API who can browse the internet in the same way and can take actions on behalf of users, as well as a more consumer -oriented offer called Mariner. Anthropic models can also perform computer tasks, including the opening of files and navigating web pages.

According to OpenAi, the new opera form, called O3 operator, was “refined with extra safety data for computer use”, including data sets designed to “teach the model” [OpenAI’s] Decision boundaries about confirmations and refusal. “

OpenAi has issued a technical report with the performance of the O3 operator on specific safety evaluations. In comparison with the GPT-4O opera model, the O3 operator will be less likely to refuse to carry out “illegal” activities and look for sensitive personal data, and less susceptible to a form of AI attack that is known as fast injection, according to the technical report.

“O3 operator uses the same multi-layered safety approach that we used for the 4O version of the operator,” Openai wrote in his blog post. “Although the O3 operator inherits the coding options of O3, it does not have any native access to a coding environment or connection.”