For the last few years, using an AI chatbot meant a fairly simple loop: you asked a question, it wrote an answer, and you took that answer away to actually do something with it. That loop has just changed. OpenAI has rolled out a new agent inside ChatGPT, aimed squarely at replacing hours of manual document, spreadsheet, and presentation work with a single instruction — describe the outcome you want, and let the agent handle the rest.
This isn't a minor feature update. It's part of a broader shift happening across the AI industry right now, with major labs racing to move their assistants out of the chat window and into the position of an actual coworker that can plan, execute, and deliver finished work with minimal supervision. Here's what's actually new, how it works, and what it means if you use ChatGPT for work.
Table of Contents
What Actually Changed
The core shift is in what the agent is trusted to finish on its own. Instead of answering a question and stopping, the new agent takes a stated goal, breaks it into a sequence of smaller tasks, and works through them independently — pulling in context from your connected apps and files along the way. Rather than handing you a block of text to copy into a document yourself, it hands you the finished document, spreadsheet, presentation, report, or even a simple website, ready to review.
It arrived alongside two related launches on the same day: a redesigned desktop app that brings chat, this new work agent, and OpenAI's coding tool into one place, and a lightweight website-building feature for turning a prompt directly into a hosted page. Together, the three releases point in the same direction — ChatGPT positioning itself less as a question-answering tool and more as a place where entire pieces of work get produced start to finish.
How the Work Agent Turns a Prompt Into a Deliverable
The mechanics are what you'd expect from an "agentic" tool, but the important detail is the persistence: once you give it a task, it can keep working in the background for hours, across devices, without you needing to keep the chat window open. You can check in, redirect it mid-task, or approve sensitive actions before they happen, but you don't have to babysit every step.
Behind the scenes, this runs on a persistent cloud-based session rather than requiring your own computer to stay on and connected — a deliberate design choice that lets a task keep progressing even after you've closed the laptop or switched to your phone.
The GPT-5.6 Model Lineup Behind It
The agent is powered by a new model family released alongside it, split into three tiers so that cost and speed can be matched to the size of the task rather than defaulting to the most expensive option every time.
Who Gets Access, and When
Access is rolling out in stages rather than to everyone at once. Pro, Enterprise, and Edu users are first in line, with Plus and Business users following within days, and broader regional availability expanding after that. If you don't see it in your account yet, it's worth checking again over the coming days rather than assuming it isn't coming to your plan.
Why This Is Happening Now
This launch doesn't exist in isolation. Anthropic's Claude Cowork, an agentic tool built for planning and executing multi-step knowledge work autonomously, has been expanding with task-specific plugins for functions like legal, sales, marketing, and data analysis. Microsoft has followed with its own agentic offering built into its Copilot ecosystem. Viewed together, these releases mark a fairly clear industry-wide pivot: the competitive battleground for AI companies is moving from "who has the best chatbot" to "whose agent can be trusted with real, unsupervised work."
That competition is also showing up in pricing and efficiency claims. OpenAI has emphasized that its smaller new models can complete tasks nearly as well as its largest ones at a fraction of the cost — a signal that the next phase of this race may be fought as much on price-per-completed-task as on raw capability.
What You Can Actually Ask It to Do
Rather than a single narrow use case, the agent is designed to be general-purpose across common categories of desk work. Some practical examples of the kind of instruction it's built to handle:
- "Pull last quarter's numbers from our shared files and put together a board-ready summary report."
- "Draft a client proposal deck based on the notes from our last three calls."
- "Reconcile this spreadsheet against the latest export and flag any mismatches."
- "Research competitor pricing for this category and build a comparison table."
- "Put together a simple landing page for this event and get it ready to share."
In each case, the distinguishing feature isn't that the AI can produce text about the topic — it's that it gathers the actual source material itself, works through the steps, and hands back a finished artifact rather than a draft you still need to assemble.
Things Worth Watching Before You Rely on It
A few practical cautions are worth keeping in mind if you're planning to hand over real work to any agent like this, regardless of which company builds it:
- Review before you send. An agent working independently for hours can drift from what you actually meant — treat the first finished output as a draft to check, not a final deliverable to forward untouched.
- Watch what it reads. If the agent uses its own browser to gather source material for a report, the quality and bias of what it finds shapes the quality of what you get back.
- Mind connected access. Linking email, calendars, or code repositories gives the agent real reach into your accounts — it's worth understanding exactly what permissions you're granting before you connect anything sensitive.
- Costs can scale with task length. A task that runs for hours in the background may consume meaningfully more usage than a single chat reply, so keep an eye on how your plan's usage is metered.
Bottom line: this generation of agents is genuinely more capable of finishing real work unsupervised than anything before it, but "unsupervised" doesn't mean "unreviewed." Treat it as a fast, capable first-pass collaborator rather than a replacement for a final human check.
Frequently Asked Questions
Is this different from the earlier "ChatGPT Agent" feature?
Yes. Earlier agent features focused on browsing the web and executing individual tasks like booking or research. This newer release consolidates that work into a broader, persistent agent aimed specifically at producing finished project deliverables — documents, spreadsheets, presentations, reports, and simple websites — rather than one-off actions.
Do I need a paid plan to use it?
Yes. It's rolling out to paid tiers first — Pro, Enterprise, and Edu — with Plus and Business access following shortly after. Free-tier access has not been part of the initial rollout.
Can it work while my computer is off?
Largely, yes. Because the task runs in a cloud-based session rather than on your local machine, a task can keep progressing even if you close your laptop, and you can check back in from your phone.
Does it replace the need to check the final work?
No. It's designed to hand you a finished-looking deliverable, but that doesn't guarantee it's error-free or exactly what you intended — a human review step before sending anything onward is still the responsible approach.
How does this compare to Claude Cowork or Microsoft's Copilot agent?
All three are built around the same core idea — an agent that plans and executes multi-step knowledge work with minimal supervision — but they differ in ecosystem, connected-app support, and pricing model. Which one fits best will likely come down to which apps and files you already work in day to day.
Whichever platform you use, the direction of travel here is clear: AI tools are moving from answering your questions to finishing your projects. That's a genuinely useful shift for busy schedules, but it also raises the bar for how carefully we review what an AI hands back before it goes out the door under our name.
Discussion