Join our daily and weekly newsletter for the latest updates and exclusive content on industry-leading AI coverage. learn more
It was a big week for AI announcements after the Microsoft, Google and Anthropic event. But Openai is using its own news to get things done. No, we’re not only talking about its $6.5 billion acquisition of Jony Ive’s design team to lead OpenAi’s new hardware effort “IO”.
Today, the company upgrades its operator-autonomous web browsing and cursor control agents from upgrading it to a new, more powerful O3 inference model using the previous GPT-4O multi-model model.
The update, released globally today on May 23, 2025, is available as a “Research Preview” to pay OpenAI’s $200 monthly Chatgpt Pro plan subscribers.
Basically, this is the way Openai says it’s not yet completely “sanded” or a perfect product – it may still have tangles and problems.
But rival Google offers its own top AI subscription bundle, which will cost close to $250 (currently enjoying discounts to $125 in the first three months) to access its latest Gemini Multimopal, Imagen Image Generation and VEO video generation models, and suddenly, Openai’s Cantgpt Pro Pro Plan seems more affordable.
What is the operator of Openai and what is it for?
The operator made its debut in January 2025, the first step for OpenAI to move towards semi-autonomous agents, especially computers using agents (CUAS). The idea is to go beyond Chatgpt’s chatbot interface and allow OpenAI’s powerful AI model to start taking more actions on behalf of users.
Therefore, operators are designed to automatically point, click, scroll and type to complete web-based tasks such as booking dinner reservations, compiling shopping lists, or ordering event tickets. This proxy function allows it to directly complete user tasks through the browser interface, from booking to collecting online data.
For security, privacy and security purposes, the operator does not use any existing web browser on the user’s PC or Mac. Instead, it can access a cloud-hosted virtual browser through a standalone website (operator.chatgpt.com), where users can enter requests in real time and watch the agent perform tasks.
It combines GPT-4O-based vision, reasoning and interaction capabilities, marking a new direction for OpenAI in AgentIC AI.
The product was released as a research preview for Chatgpt Pro subscribers and incorporates built-in security measures such as user confirmation, watch mode, and restrictions on high-risk network platforms.
It has also been tested in an enterprise environment, including travel planning and civic services, demonstrating its potential in both consumer and business environments.
O3 provides improved accuracy, structure and success rate
With this update, OpenAI aims to improve performance in several key dimensions. The new O3-based operator demonstrates improved durability and accuracy during browser interactions.
In fact, this means it is more likely to successfully complete user tasks and requires less correction or repetition. Additionally, users can expect a clearer, more structured and more comprehensive response.
In comparative evaluation, the new model has a clear preference advantage over its predecessor. Human preference studies show that users prefer the style, comprehensiveness and clarity of the O3 model. It also performed well in compliance and efficiency, although the factual correctness results between the versions are more balanced.
The performance of third-party evaluation benchmarks reflects these enhancements. In the OSWORLD benchmark that measures browser-based tasks, the O3 model scored 42.9, compared with the previous version of 38.1.
However, OpenAI pointed out that due to the limitations of the automated grading system, the actual performance gain may be close to 20 percentage points!
On Webarena, the new model scored 62.9, up from 48.1. The most compelling improvements appeared in the Gaia benchmark, with the O3 model scored 62.2, surpassing the previous model’s 12.3.
Side by side comparison further illustrates these benefits. In an example involving restaurant booking requests, the new model offers a clearer, more detailed list of available bookings, including locations, Michelin ratings and seat notes, presented on a well-formed table. According to the new O3 Operator Release Notes: Previous versions provided less information in less functional ways:
Safeguards remain, like general warning notes for sensitive financial transactions and account access
The O3 model also inherits security measures introduced in earlier versions and further fine-tunes its role as a proxy system.
OpenAI has integrated enhanced training to prevent harmful task execution, timely injection vulnerabilities, and errors involving user intent.
The evaluation shows that the model now confirms 94% of the sensitivity measures before implementing them and 100% confirms in financial transactions. Rapid injection susceptibility also decreased from 23% to 20%.
It is worth noting that O3 operators maintain cautious boundaries on certain high-risk web interactions, such as email or financial platforms, where user supervision may be required through watch mode or explicitly refuse to proceed. These measures are part of a layered security approach that combines model-level robustness with real-time monitoring.
While the upgrade to the operator marks an improvement in technology, it also reflects OpenAI’s ongoing commitment to responsible for AI deployment.
The system’s ability to take practical actions introduces new risks and the development team continues to refine its security protocols accordingly.
According to OpenAI’s updated O3 system card documentation, the model maintains a low risk capability threshold in categories such as biology and chemical abuse, and has no local coding environment or terminal access, further reducing potential abuse vectors.
The operator is still a research preview, only for Chatgpt Pro users. At least for now, operators responding to API versions will continue to be based on the GPT-4O model.
Impact on corporate technical decision makers
Upgraded operators will significantly enhance the workflow of AI engineering, orchestration, data management and IT security professionals.
For those who build or maintain a machine learning model, the improved accuracy and structured output of the model reduces the overhead of test validation and troubleshooting.
In an orchestration environment, it provides a practical, reliable tool for automating browser-based components of complex pipelines.
Data engineers can delegate manual web interactions such as data verification and scratching and provide more confidence to provide free time for advanced optimization efforts.
Meanwhile, due to the model’s hierarchical security mechanism, security professionals have obtained a safer approach to simulate user behavior in audit and incident response exercises.
In these disciplines, O3-based operators both introduce functional upgrades and risk mitigation frameworks, making it a practical addition to modern technology toolkits.
