ChatGPT Work Explained: Long-Running Jobs Inside ChatGPT

ChatGPT began as a place you opened when you needed an answer. ChatGPT Work pushes it further: instead of replying and stopping, it takes on long-running jobs across your apps and files and keeps going for hours. If ordinary chat is asking a colleague a question, Work is handing that colleague a project folder and coming back at the end of the day.

This guide explains what Work is, how it differs from both chat and the newer ChatGPT Dots, what to hand it, and where the honest limits sit. Some details are still emerging, and where that is true, this article says so rather than filling the gaps with guesses.

Key takeaways

• ChatGPT Work handles multi-hour assignments across apps and files, going well beyond single-reply chat.
• Tasks started or managed through Codex or ChatGPT Work count toward those services' usage limits.
• Work finishes defined jobs; Dots add persistent background monitoring and cross-channel continuity on top.
• Good Work jobs look like contractor briefs: clear inputs, clear outputs, and checkable intermediate steps.
• Access details for Work are less published than the Dots rollout, so verify current availability on OpenAI's pages before planning around it.

What ChatGPT Work actually is

The core idea is simple. You describe a complex assignment, point at the relevant apps and files, and Work carries it forward over hours rather than seconds. It can move between tools, consult documents, and assemble results while you do something else. Reporting on OpenAI's agent launches describes Work as already able to spend hours completing complex assignments once they are handed over.

Two terms need untangling because OpenAI uses them in overlapping ways. "Chat" is the familiar back-and-forth: prompt, answer, follow-up. "Work" is delegation with duration: the system accepts a job, works through it across your connected context, and returns with something closer to a deliverable than a reply. Codex enters the picture as one of the surfaces where such jobs can be started and managed, and usage accounting follows the service: tasks run through Codex or Work count toward those services' limits rather than being free.

For a business owner, the mental model that works best is the contractor brief. If you could hand the job to a capable freelancer with a folder of inputs and a definition of done, it is shaped like a Work job. If it needs an instant answer, use chat. If it needs ongoing watching rather than a finish line, that is the territory of Dots, covered next.

How Work differs from chat and Dots

These three get blurred in marketing, so here is the practical split.

The key sentence from launch reporting: Dots add ongoing background monitoring and cross-channel continuity on top of what Work already does. Work completes the assignment; Dots stick around afterward. Our ChatGPT Dots explainer covers the always-on side in full, and the Dev Day 2026 recap puts both in the context of OpenAI's announcements.

One honest note: OpenAI has documented Dots, their availability, and their controls in considerable detail, while Work's standalone specifics remain thinner in public sources. That asymmetry is worth knowing because it tells you which of the two is easier to plan a business process around today.

What to hand it: jobs that fit

The jobs that fit Work share a shape. They take a person hours, they span more than one file or app, the inputs are available up front, and the output can be checked against something. Examples in that shape include compiling a research pack from a folder of sources, reconciling two versions of a report into one clean draft, turning interview transcripts into structured social posts, re-running an analysis when the underlying numbers change, and preparing a first draft of a proposal from templates and past bids.

Each of those has a finish line and a reviewer. That pairing matters more than the choice of tool. A Work job without a named reviewer becomes output nobody trusts; a Work job without a finish line becomes an open-ended ramble that burns usage limits. Tasks run through these services count toward their limits, so an unbounded job is not just vague, it is metered vagueness.

Write the brief the way you would brief a contractor. State the goal in one sentence, list the inputs and where they live, name the steps, define what done looks like, and say who approves the result. If you cannot write that brief, the job is not ready to delegate to software, and handing it over anyway just moves the confusion downstream. The internal assistant service page shows how structured briefs and approvals look inside a working deployment.

What not to hand it

Some jobs are wrong for Work regardless of how good the underlying model is. Anything that contacts customers unsupervised, sends money, changes credentials, or alters records of account should wait for a setup with explicit approval tiers, which is the Dots pattern with its allow, block, and require-approval rules, not a fire-and-forget job. Anything regulated, legal filings, medical documentation, financial reporting, needs a human professional in the loop as a matter of policy, not preference.

Also keep Work away from jobs where the inputs are a mess. Agents amplify the state of your data: clean folders produce clean output, and a drive full of duplicates and outdated versions produces confident confusion. If your files are disorganized, the highest-return move is tidying the inputs first, ideally with help from structured internal assistant workflows rather than another tool layered on top.

Limits and unknowns, stated plainly

A new capability deserves a clear-eyed limits section, especially when public documentation is still catching up.

First, usage accounting: Work and Codex tasks count toward those services' limits, so long jobs are metered. Get the limit terms in writing for your plan before running a week of multi-hour jobs, and watch the first invoice the way you would watch any new vendor bill.

Second, availability is a moving target. The Dots rollout has published specifics, Pro and Business Premium in eligible markets, Enterprise beta behind an admin switch, regional exceptions, while Work's access story is less crisply documented. Confirm the current state on OpenAI's own pages rather than relying on any article, including this one, for access planning.

Third, accuracy compounds over long runs. A small misunderstanding in hour one becomes a large wrong deliverable in hour four. The defense is checkable milestones: ask for intermediate outputs at defined points so a drift gets caught early. This is the same discipline as managing a human contractor, and it works for the same reason.

Fourth, model specifics: Dots run on GPT-6 Astra, and detailed independent benchmarks for that model were not yet published at the time of writing. Judge the system by what your pilot produces, not by launch-week adjectives.

A safe way to try it

Run a two-week trial on one job. Pick a recurring assignment with a clear finish line, write the one-page brief, name a reviewer, and run it twice through Work while a person does it the old way in parallel. Compare the outputs on correctness, completeness, and time spent reviewing. Track usage against your plan limits so the cost side is measured, not assumed.

If the trial output needs heavy correction, fix the brief before blaming the tool: add examples of good past output, narrow the sources it may use, and insert a midpoint check. If the trial works, expand to adjacent jobs in the same function before spreading to new teams. And if your comparison shopping includes Anthropic, our Claude side guide and the Claude vs ChatGPT business comparison give you the other half of the picture.

FAQ

What is ChatGPT Work?

ChatGPT Work is an OpenAI capability for long-running jobs that span apps and files. Instead of answering in one reply, it keeps going for hours once you hand over a complex assignment. Tasks started or managed through Codex or ChatGPT Work count toward those services' usage limits.

How is ChatGPT Work different from ChatGPT Dots?

Work is built for finishing a defined assignment that takes hours, while Dots are persistent agents that keep monitoring your goals in the background across ChatGPT, Slack, Teams, and voice. Reporting on the launch describes Dots as adding ongoing background monitoring and cross-channel continuity on top of what Work already does.

Who can access ChatGPT Work?

OpenAI has not published a standalone access page for Work with the clarity of the Dots rollout, which covers Pro and Business Premium in eligible markets plus an admin-gated Enterprise beta. Treat Work availability as evolving, confirm current access on OpenAI's own pages before planning around it, and keep claims about it modest.

What kinds of jobs suit ChatGPT Work?

Multi-hour assignments with clear inputs and checkable outputs: compiling research across files, reconciling reports, preparing draft deliverables, and multi-step analyses. It suits work you could brief to a capable contractor, not instant answers, not unsupervised customer contact, and not anything where the approval path is unclear.

Next step

Pick one recurring multi-hour job, write the one-page brief, and trial it for two weeks against the manual process with usage tracked. If you want help choosing the job and setting up the review loop, book a call or start with the free six-step AI automation plan on our homepage.