Need smarter insights in your inbox? Join our weekly newsletters to get solely what issues to enterprise AI, information, and safety leaders. Subscribe Now
OpenAI isn’t letting the delay of its open supply AI mannequin gradual it down on delivery different options.
At this time, the firm is unveiling ChatGPT agent, a characteristic that permits its AI chatbot to autonomously browse the net, conduct in depth analysis, obtain and create new information for its human customers utilizing its personal digital pc.
Come once more? ChatGPT now will get its personal PC? And it may possibly use that PC to log into your, the human consumer’s, accounts and obtain or ship stuff for you?
That’s appropriate, not less than in a digital sense, in response to OpenAI. As the corporate explains:
The AI Influence Sequence Returns to San Francisco – August 5
The following section of AI is right here – are you prepared? Be part of leaders from Block, GSK, and SAP for an unique have a look at how autonomous brokers are reshaping enterprise workflows – from real-time decision-making to end-to-end automation.
Safe your spot now – area is restricted: https://bit.ly/3GuuPLF
“The mannequin can select to open a web page utilizing the textual content browser or visible browser, obtain a file from the net, manipulate it by working a command within the terminal, after which view the output again within the visible browser. The mannequin adapts its strategy to hold out duties with pace, accuracy, and efficiency.”
The best way to use ChatGPT agent
Customers can interact the agent by clicking on the ‘Instruments’ button within the ChatGPT immediate entry field, opening the dropdown menu, and choosing ‘agent mode’ from the accessible choices.
Then, when it’s turned on describe a job in plain language, and the agent can carry it out throughout internet and native app environments, combining reasoning with actions that beforehand solely a human consumer might carry out on their very own machine, manually.
ChatGPT agent can hook up with apps like your private or enterprise Gmail and GitHub, so it may possibly pull in helpful data — emails or code — out of your accounts to assist with duties you ask it to do. It might probably hook up with third-party software programming interfaces (API) to drag data and use linked functions and companies by way of them, as effectively.
If an internet site wants you to log in, you are able to do that securely by way of a particular browser view, which lets the agent dig deeper and deal with extra personalised duties, like checking your inbox or filling out kinds in your behalf.
Going the place Operator couldn’t — offline
The brand new ChatGPT agent builds upon and expands from the “Operator” agent OpenAI launched in January 2025, which allowed ChatGPT to browse the net and fill out kinds, place orders, and do different web-based duties in a non-public “headless browser,” that’s, a cloud-based customized internet browser that OpenAI itself maintained and supplied for every Operator session.
Nonetheless, Operator was restricted solely to interacting with web sites and web-based functions — not packages that would additionally run regionally on a PC, similar to spreadsheet tabulators and slide deck presentation making software program.
The brand new ChatGPT agent can browse web sites, work together with on-line kinds, run code, analyze information, and ship completed outputs — similar to editable shows or spreadsheets — primarily based totally on consumer directions.
The disclosing comes on the heels of a report printed days in the past by unbiased subscription tech business web site The Info suggesting that OpenAI would improve ChatGPT to be a extra direct competitor to its personal investor Microsoft’s Workplace software program functions (e.g. Excel, Phrase, PowerPoint, and many others.)
Merging Operator and Deep Analysis into one agent
In truth, OpenAI positions ChatGPT agent as a merging of two of its prior brokers — Operator and Deep Analysis, the latter launched in February 2025, which exhaustively searches the net by way of its personal headless text-only browser to seek out and compile data into prolonged and in-depth stories (therefore the identify). As OpenAI writes in a weblog submit:
“Operator couldn’t dive deep into evaluation or write detailed stories, and deep analysis couldn’t work together with web sites to refine outcomes or entry content material requiring consumer authentication. In truth, we noticed that many queries customers tried with Operator have been truly higher suited to deep analysis, so we introduced the perfect of each collectively.”
The earlier Operator instrument might be phased out, however customers can nonetheless entry Deep Analysis by way of the dropdown within the ChatGPT interface.
Whether or not utilizing a visible browser to work together with an internet site or a terminal to run Python code, the agent strikes seamlessly between instruments inside a single session.
It helps a broad vary of use circumstances, from analyzing opponents and producing stories to planning journeys, summarizing emails, or reserving appointments.
Customers can interrupt, redirect, or pause a job at any time, with the agent choosing up proper the place it left off.
Availability and entry
Beginning at the moment, subscribers to ChatGPT’s $200-per-month “Professional” tier can have full entry to ChatGPT agent, with a month-to-month quota of 400 messages.
ChatGPT Plus ($20 per thirty days) and Group ($30 per thirty days) will acquire entry over the subsequent few days, with 40 messages per thirty days. Further utilization is offered by way of credit-based choices.
OpenAI mentioned in a launch shared with VentureBeat underneath embargo that its ChatGPT Enterprise and Schooling subscribers will acquire entry to the characteristic the approaching weeks.
For now, the characteristic isn’t but accessible in Europe or Switzerland, little question disappointing residents there.
Constructed with security and management on the forefront
On condition that the agent can now take actions on behalf of customers — together with on logged-in web sites or with linked apps — OpenAI has launched in depth security measures.
These embody consumer confirmations earlier than taking motion, energetic supervision for delicate duties, and technical safeguards to restrict unintended conduct.
Key protections embody:
- Affirmation prompts earlier than actions like submitting kinds or sending emails
- Watch Mode, which pauses execution when a consumer turns into inactive
- Refusal of high-risk duties, similar to monetary transfers or privateness violations
- No reminiscence retention throughout agent classes
Excessive-risk area classification
According to its Preparedness Framework, OpenAI is treating ChatGPT agent as a Excessive functionality system within the organic and chemical domains.
Whereas there isn’t a direct proof of misuse, the corporate is activating its strongest security safeguards out of warning.
These embody enhanced refusal coaching, pink teaming by biosafety specialists, and improved detection programs.
Recall that Anthropic‘s latest data launched about its new Claude Opus 4, and different surveys of superior AI fashions, have proven that when given entry to exterior instruments and functions similar to electronic mail, they will in some circumstances take actions they imagine to be ethical and moral however which will compromise the consumer, similar to emailing authorities companies or journalists of suspected wrongdoing on the a part of the consumer.
The mannequin thinks it’s appearing like a “whistleblower” however in truth, could compromise consumer privateness, safety, and proprietary data and alert authorities to wrongdoing the place there’s none, or the place it’s doubtful.
Robust efficiency on real-world duties
ChatGPT agent isn’t simply extra succesful in concept—it has delivered robust outcomes throughout quite a few benchmarks designed to simulate real-world information work. It set a brand new excessive rating on Humanity’s Final Examination with a 44.4 utilizing parallel rollout strategies, and achieved 27.4% on the troublesome FrontierMath benchmark.
On SpreadsheetBench, it scored 45.5%—greater than doubling Copilot in Excel’s efficiency.
Present limitations and subsequent steps
Some options, like slideshow technology, are nonetheless in beta and should really feel primary in formatting or differ barely between in-app previews and exported information.
OpenAI is actively coaching the subsequent iteration of this characteristic to enhance polish and format.
The launch of ChatGPT agent indicators a shift in how customers work together with AI—from asking inquiries to assigning full duties.
With its means to motive, act, and produce deliverables, OpenAI is betting that customers more and more need AI not simply to help them, however to work for them. Whereas the corporate emphasizes that the agent continues to be evolving, it sees this launch as the muse for a extra interactive, action-oriented future for AI.
