It was Holy Week for AI announcements after the events of Microsoft, Google and Anthropic. But Opeli ends things with its own news. And no, we do not only talk About $ 6.5 billion to lead New hardware effort, “IO” at Openai.
Today The company updated the operator Autonomous browsing and agent cursor at CHATGPT from using the previous multimodal model of a large GPT-4O language to a newer and stronger model of O3 reasoning.
The update, issued globally today, on May 23, 2025, is available as a “research preview” for paying subscribers of the ChatGPT Pro plan value USD 200 in the amount of USD 200.
Basically, this is the way OpenAi said that it is not yet fully “polished” or an excellent product – it should still have breakdowns and problems.
But from Google’s rival offers its own top -level AI subscription package for a price of just about USD 250 commonly (Currently discounts up to USD 125 in the first three months) to access the latest Multimodal Gemini, Imagen Image Generation and VEO Video Generation models, suddenly the ChatgPT Pro OpenAI plan seems more accessible.
What is the Openai operator and what is it for?
The operator debuted for the first time in January 2025 as the initial OpenAI step to semi -automatic agents, especially computer ones using agents (Cuas). The idea is to go beyond the chatbot chatgpt interface and let the powerful AI Openai models start more actions on behalf of the user.
In this fashion, the operator was designed for an autonomous point, clicking, changing and writing to perform web tasks reminiscent of booking dinner reservations, compilation of shopping lists or ordering tickets for events. This agency ability allows him to perform the user’s tasks directly via the browser interface, from booking booking to online data collection.
For safety, privacy and safety purposes, the operator didn’t use any existing web browser on a user or Mac. Instead, he operated in a virtual browser in a cloud available through an independent website-lator.chatgpt.com-where users could introduce demands and observe the agent of real-time tasks.
He combined the possibilities of vision, reasoning and interaction based on GPT-4O, marking a latest direction for OpenAI at Agentic AI.
The product was launched as a research preview for CHATGPT PRO subscribers and contained built -in security measures, reminiscent of user confirmation, viewing mode and limitations on high -risk web platforms.
He was also tested in the context of enterprises, including in planning travel and civic services, showing its potential in each consumer and business environments.
O3 offers higher accuracy, structure and success indicators
Thanks to this update, OPENAI is aimed at increasing performance in several key dimensions. The latest O3 operator shows higher durability and accuracy during the interaction of the browser.
In practice, because of this it is more likely that he successfully performs the user’s tasks and with a smaller need to correct or repeat. In addition, users can expect answers that are clearer, more structured and more comprehensive.
In comparative assessments, the latest model shows a clear advantage of preferences in relation to its predecessor. Studies of human preferences reveal that users favor the O3 model due to its style, comprehensiveness and clarity. It also results strongly in terms of instructions and performance, although the results for the actual correctness are more balanced between versions.
The performance of comparative tests about the assessment of one other company reflects these improvements. On Osworld benchmark This measures the completion of the tasks based on the browser, the O3 model results 42.9 compared to 38.1 for the previous version.
However, OpenAI notes that due to restrictions in the automated assessment system, the actual performance increase might be closer to 20 percentage points!
In Webarena, the latest model achieved a results of 62.9, compared to 48.1. The most dramatic improvement appears in Gaia, in which the O3 model is 62.2, significantly exceeding the previous model 12.3.
Comparisons of tasks next to each other moreover illustrate these profits. In one example, the latest model provided a more pronounced and more detailed list of accessible reservations, including locations, Michelin rankings and notes from the seats, presented in a well -formated table. The previous version, although functional, provided less information in a less organized way, according to the image contained with New comments of the OPO operator release:
Security stays, in addition to general warning remarks regarding the use of sensitive, financial transactions and access to the account
The O3 model also inherits security measures introduced with previous versions, and further tuning for its role of the agency system.
Opeli integrated improved training with harmful performance of tasks, fast injection gaps and errors related to the user’s intentions.
The assessments show that the model now confirms 94% of sensitive actions before their implementation, with 100% confirmation of monetary transactions. The rapid injection susceptibility also decreased from 23% to 20%.
In particular, the O3 operator maintains a cautious limit of some high-risk web interactions, reminiscent of e-mail or financial platforms, in which it might require user supervision through the statement mode or clearly refuse to proceed. These measures are a part of the layered security approach that mixes the reliability of the model level with real -time monitoring.
While the operator’s update means technical improvement, it also reflects OpenAI’s continuous involvement in the responsible implementation of AI.
The system’s ability to take real actions introduces a latest risk, and a team of programmers continues to properly decrease their security protocols.
According to Updated documentation of the OPENAI system cardThe model stays below the thresholds of high risk capability in terms of reminiscent of improper biological and chemical use and there is no native coding environment or terminal access, which moreover reducing potential vectors of improper use.
The operator stays a view of the research and is only available to ChatgPT Pro users. The API API version will proceed to be based on the GPT-4O model, at least for now.
Implications for technical decision makers of enterprises
An improved operator can significantly increase the flows of pros in the field of AI engineering, orchestration, data management and IT security.
For those that build or maintain machine learning models, improved accuracy and structured model results reduce the costs of validation of testing and problem solving.
In the context of orchestration, it offers a practical, reliable tool for automating components of complex pipelines based on browser.
Data engineers can delegate manual web interactions-such as data verification and scraping-with greater trust, releasing time for optimization work at a higher level.
Meanwhile, safety specialists gain a safer way of simulation of user behavior in audits and reacting exercises to incidents, thanks to the layered model safety mechanisms.
In these disciplines, the O3 operator introduces each the improvement of the ability and the framework of risk reduction, which makes it a practical addition to a modern technical set of tools.
