Join our daily and weekly newsletters for the latest updates and special content related to the industry’s leading AI coverage. Get more information
It was a big week for AI announcements after the Microsoft, Google, and Anthropic incidents. But Openi is eliminating things with his news. And no, we’re just not talking Jonny Ive’s Design Team About Acquisition of $ 6.5 Billion To guide a The new effort of the hardware, “io” in Openai.
Today, The company upgraded its operator Chat GPT’s autonomous web browsing and cursor controlling agent new and more powerful O3 reasoning model by using the former GPT-4O multi-modal large language model.
This refreshment, which is released globally on today, May 23, 2025, is available as an “research preview” for the payment of subscribers of the Openi US $ 200 Monthly Chat GPT Pro Plan.
Basically, it is an open way to say that it is not yet a completely “sanded” or complete product – it can still have concrete and problems.
But with Rival Google regularly offers its Top Tero AI subscript bundle for about $ 250 US $ 250 –
What is Openi operator and what’s for it?
The operator first debuted in January 2025 as an initial step in the semi -autonomous agents of the Openi, especially the computer using the Cuas. The idea is to move beyond the chat boot interface with a chat boot interface and start taking further steps by the user to the powerful AI model.
Thus, the operator was designed to click, scroll and type, with sovereignty to complete web -based tasks such as booking of dinner protection, compiling purchases, or ordering event tickets. This agent capacity allows it to complete the user tasks directly through the browser interface, from booking reservation to collecting data online.
For safety, privacy and security purposes, the operator did not use any existing web browser on the user’s PC or Mac. Instead, it is a standstone site-operator. Chat went to a cloud -hosted virtual browser by Chat GPT.com where users can input applications and act in real time to the agent.
It combined the GPT -4O -based vision, reasoning and interaction capabilities, which indicate a new direction for the open in the agent AI.
This product was launched as a research preview for Chat GPT Pro users and includes restrictions on built -in safety measures such as user verification, watch mode, and high -risk web platforms.
It was also being tested in the enterprise context, including travel planning and urban services, which showed its ability in both consumers and the business environment.
O3 offers better accuracy, structure and success rate
With this update, the openness aims to increase performance in many important dimensions. The new O3 -based operator shows better resilience and accuracy during the browser conversation.
In practical terms, this means that consumer tasks are more likely to complete successfully and with less need of correction or repetition. In addition, consumers can expect answers that are clear, more structural and more comprehensive.
In a comparative diagnosis, the new model shows more preferred benefits from its predecessor. Human priority studies suggest that users are in favor of the O3 model for their style, comprehensiveness and explanation. It also performs strictly in the following and performance recipe, though the results of the accuracy of the facts are more balanced among the version.
The performance of the third party’s diagnostic benchmark reflects the increase. On Os World Benchmark Mark It measures the completion of browser -based tasks, the O3 model scored 42.9 compared to 38.1 of the previous version.
However, Openi notes that due to limits in the automatic grading system, the actual performance benefit can be close to 20 % points!
On Webrina, the new model scored more than 48.1 62.9. The most dramatic improvement appears on the GAIA benchmark, where the O3 model scores 62.2, which goes ahead to 12.3 of the previous model.
The comparison of the work in the side makes these benefits more clear. An example, containing a restaurant booking request, the new model provided a clear and more detailed list of available reservations, including locations, Michalain ratings, and seating notes, which are presented in a well -manufactured table. The previous version, while actively, delivered less information in a less systematic manner, according to a picture included New O3 Operator Release Note:
Safe Guards are left, such as general precautions on use on sensitive, financial transactions and account access
The O3 Model also inherited the safety measures introduced with the previous version, which, as an agent system, has a more good toning for its role.
The Open AI has integrated better training against the implementation of the harmful task, the immediate injection risks, and the consumer intended mistakes.
The diagnosis shows that the model now confirms 94 % sensitive measures before implementing them, which contains 100 % confirmation in financial transactions. Injection’s sensitivity has also been reduced from 23 % to 20 %.
In particular, the O3 operator maintains a cautious limit on some high -risk web interactions, such as email or financial platforms, where it may need user monitoring through watch mode or refuse to move forward. These steps are a part of a one -layered approach to safety, which connects the model level strength with real -time monitoring.
Although upgrade to the operator indicates technical improvement, it also reflects on the ongoing commitment to the opening AI responsible AI.
The real -world measures system introduces new risks, and the development team continues to improve its safety protocol accordingly.
According to Openai’s latest O3 System Card DocumentsThe model is below the threshold of biological and chemical abuse such as high risk capacity in category and has no coding access to a local environment or terminal, which can further reduce potential misuse vector.
The operator is a research preview and is only accessible to Chat GPT Pro users. The operator’s answers will be based on the GPT-4O model, at least for now.
Improvements for Enterprise Technical Decision Makers
The upgraded operator has to significantly increase the workflows of professionals in AI engineering, orchestration, data management, and IT security.
Machine learning models reduce the overhead of construction or maintenances of the model, better accuracy of the model and structural output test verification and defects.
In the orchestration context, it offers a practical, reliable tool to automate the browser -based components of complex pipelines.
Data engineer manual web interactions-such as data verification and scraping-more confidence, can free the time for high-level correction work.
Meanwhile, security professionals get a safe way to imitate user behaviors in audit and event response exercises thanks to the model layered protective procedures.
In these articles, the O3 -based operator introduces both the capabilities and a risk reduction framework, which enhances the modern technical tool cut.
