Openai General Director, Altman himself began this year, saying in the post on the blog that 2025 Be big for AI agentsTools that can automate tasks and take action on your behalf.
Now we see the first real test of OPENAI.
Opeli announced on Thursday that he launches the preview of the operator’s research, the general AI agent who can take control of the web browser and independently perform specific actions. The operator comes to US users in a pro pro subscription CHATGPT in the amount of USD 200. Opeli claims that it is finally planning to introduce this function of more users at the plus, team and enterprise level.
“[Operator] will be [in] Soon other countries – said Altman himself, the general director of OPENAI on Thursday live. “Unfortunately, Europe will take some time.”
This initial research preview is available via the operator.chatgpt.com, but soon OpenAI claims that he wants to integrate the operator with all ChatgPT clients.
The operator promises to automate tasks such as travel accommodation booking, booking restaurants and online shopping, according to OPENAI. There are several categories of tasks from which users can choose in the operator interface, including purchases, delivery, restaurant and travel – all allow different types of automation.
When CHATGPT users activate the operator, a miniature window will appear showing a dedicated web browser, which the agent uses to perform tasks, along with the explanations of specific activities that the agent performs. Users can continue to take control of their screen while the operator is working because the operator uses its own dedicated browser.
Opeli claims that the operator is powered by a computer agent or CUA model, which combines the capabilities of the vision of the GPT-4O model of the company with the ability to reason with more advanced OPENAI models. CUA is trained to interact with the front of websites, which means that it does not have to apply API interfaces addressed to programmers to apply various services.
In other words, CUA can apply buttons, move the menu and fill out the forms on the website, just like man.
Opeli claims that he cooperates with companies such as Dordash, Ebay, Instacart, Priceline, Stubhub and Uber to make sure that the operator respects the services of services of these companies.
“The CUA model is trained to ask for confirmation of the user before finalizing tasks with external side effects, for example, before placing the order, sending e-mails, etc. so that the user can check the model’s work twice before it becomes permanent” Openai writes in materials delivered to TechCrunch. “[It] It has already proved to be useful in various cases and we try to expand this reliability in a broader extent of tasks. “
But OpenAi warns that CUA is not perfect. The company says “[doesn’t] to expect [the] CUA for reliable performance in all scenarios. “
“Currently, the operator cannot reliably support many complex or specialized tasks,” OpenAI he adds In the “such as creating slide shows, management of convoluted calendar systems or interaction with highly adapted or non -standard interfaces.
OpenAi also requires supervision of some tasks, such as banking transactions, CUA and the operator, can act mainly on their own. For example, users will have to provide information about the credit card. Opeli claims that the operator does not collect or screenshot any data.
“On particularly sensitive websites, such as E -Mail, the operator requires active user supervision, ensuring that users can directly catch and solve any potential errors that the model can make,” says OpenAI in their support materials.
This limits the usefulness of the operator – but also ensures that the agent does not hallucinate and, say, issuing a mortgage repayment on the accent chairs. Google has adopted a similar approach to his Mariner AI Agent project, which also does not complete information such as credit card numbers.
Limitations
The operator has several restrictions that are worth paying attention to.
There are limits-everyday and dependent on the task. Opeli says that the operator can perform many tasks at the same time, but there are “dynamic boundaries”. There is also a general apply limit that reset every day.
At this stage, the operator will also refuse to perform tasks for security reasons, such as sending E -Maili (despite the fact that CUA is capable of it) and removing calendar events. Opeli says that this will change in the future, but it doesn’t give ETA.
The operator can also “get stuck” if it runs in a particularly convoluted interface, password field or Captch check. He will ask the user to take over, when this happens, says Opeli.
Aggressive future
Opeli was rather tardy, developing an AI agent compared to rivals (see agents from Rabbit, Google and Anthropic), which may have something to do with safety threats around technology.
When the AI system can take action on the internet, it opens the door to much more hazardous cases of apply than vile actors. You can automate AI agents to organize phishing fraud or DDOS attacks or have their tickets for the concert before anyone else. Especially in the case of a tools as widely used as chatgpt, it is crucial that Openai takes steps to prevent such feats.
It seems that Opeli believes that the operator is sheltered enough to publish in its current form, at least as a preview of the research.
“The operator uses tools that are aimed at limiting the model’s susceptibility to malicious hints, hidden instructions and phishing tests,” OpenAI explains on your website. “The monitoring system stops performing in the event of suspicious activity, while the pipelines automated and reviewed by man constantly update security.”
The operator is the most crucial OpenAI attempt in creating an AI agent. Last week, Opeli issued tasks, providing uncomplicated chatgpt automation functions, such as the ability to set reminders and prompt to start at the set time.
The tasks gave users to ChatgPT of some friends, but the necessary functions to make ChatgPT practical as Siri or Alexa. However, the operator shows the possibilities that the previous generation of virtual assistants could never do.
AI agents were put as another great thing in artificial intelligence after chatgpt: a recent technology that will change the way the Internet and their computers will be used. Instead of simply providing and processing information, agents can – theoretically – take action and actually do things.
With the release of the first specific OpenAI agents, it will soon become clear how realistic this vision is.