ChatGPT Work Does Run On A Schedule. The Catch Is What Each Run Costs.

Correction, 8 September 2026. An earlier version of this article argued that ChatGPT Work cannot be scheduled, that it "builds the thing but does not remember to." That was wrong. It was written from secondary coverage, two pieces of which contradicted each other, and the disagreement was resolved in the wrong direction. OpenAI's own announcement says ChatGPT Work can run on a schedule, fire when an event occurs, and use your connected apps while doing it. The section below has been rewritten against that source, and the seven use cases have been re-grounded on named customers rather than generic examples. The rest of this site's claims are traced to primary documentation; this one was not, and it should have been.

ChatGPT Work is an agent inside ChatGPT that takes a goal and returns finished material rather than an answer you then have to rebuild. Sheets, slides, docs, and web apps. It launched on 9 July 2026, powered by GPT-5.6, with Codex technology underneath.

The framing that matters is the output. Ordinary ChatGPT gives you a response you copy into the real artifact. ChatGPT Work gives you the artifact. You follow its progress, answer its questions, redirect it, and approve the actions that matter.

Seven things it is doing in real organizations

These come from OpenAI's launch announcement, which names the customers. Worth reading in that spirit: they are vendor-supplied success stories, not independent case studies. The specifics are still more useful than a generic list, because you can check them against your own work.

1. Lead review at volume. Zapier built a repeatable system for reviewing thousands of leads a month. It traced customer touchpoints across their CRM, email and other tools, found where follow-ups broke down, and generated a weekly executive dashboard. Zapier says it surfaced "seven figures in potential sales."

2. Month-end close. OpenAI's own finance team reports going from days to hours: finding source data, moving it into Excel or Sheets, reconciling it, building slides, and verifying the results.

3. Competitive analysis you can audit. Virgin Atlantic gave it a customer journey and a list of competing airlines, and asked it to research what each offered and assess where Virgin led or lagged. The output was a dataset the team could review and refine, not a wall of prose. Weeks of analysis to hours, as input to a five-year plan.

4. Release and launch checks. RingCentral turned manual monthly launch reviews into a repeatable workflow across release plans, Jira tasks and go-to-market schedules, producing source-backed reports naming owners and next steps. Their R&D efficiency manager went from supporting one product manager to roughly fifty.

5. Event preparation and the debrief nobody does. NVIDIA replaced an Excel workflow that ate about 40% of one manager's pre-conference time, then used it afterward to synthesize hundreds of session transcripts and meeting notes. The two-week review became a discussion instead of a data-assembly exercise.

6. A discovery call turned into a proof of concept. OpenAI's sales team reports going from a conversation to a tailored proof of concept in 24 hours, a process that normally takes weeks.

7. Ship a Site. Sites, in public beta, turns work into an interactive site or web app you share by URL: live dashboards, project trackers, launch calendars, prototypes, internal portals, interactive reports. ChatGPT can update them as the underlying information changes.

The common shape across all seven: messy inputs, a known output format, and no judgment needed in the middle. That is the zone where this pays.

It does run on a schedule, and it does fire on events

This is the part most coverage got wrong, including an earlier version of this article.

OpenAI's announcement is unambiguous:

"Scheduled Tasks let you ask ChatGPT to perform an action once, repeat it on a schedule or when an event occurs, or monitor for changes over time."

And it is presented as part of ChatGPT Work, not a separate consolation prize:

"Even when you're away from your computer or phone, ChatGPT Work can keep projects moving forward with Scheduled Tasks. For example, it can independently turn new messages from Microsoft Teams and Slack into updated docs or slides, then share important changes with your team."

It is also not context-blind. OpenAI's examples have Scheduled Tasks using your connected apps and the browser: reviewing new Slack updates each week and refreshing a recurring agenda, checking dashboards each morning and sending a summary of what changed, monitoring customer feedback and turning recurring themes into prioritized product ideas, updating a presentation when new feedback arrives by email.

On desktop it goes further. Computer Use lets ChatGPT operate your machine, clicking, typing and moving files, and OpenAI states it works "for a one-time task or as part of a Scheduled Task when recurring work includes steps on your computer."

!What starts a ChatGPT Work run: once, on a schedule, when an event occurs, or monitoring for changes. What it reaches while running: connected apps, the browser, and your computer via Computer Use\n\nSo "every Friday at four" exists. "When a deal closes" exists, to the extent your CRM emits something a monitor can see.

The real constraint is the meter

If capability is not the limit, what is?

Cost, and the fact that nobody has published what a run costs.

OpenAI is direct that this bills differently from chat:

"ChatGPT Work is designed for longer, more involved work than a typical chat request, so usage works differently. Usage varies with the amount of work required, and more complex tasks may use more of your plan's included usage. ChatGPT Work follows the same usage structure as Codex."

There is no published per-task rate. A task's cost depends on how much work it turns out to need, which you do not know before you run it, and which a scheduled task will incur again on every firing.

The strongest evidence that this matters is what OpenAI shipped alongside it. Enterprise and Edu admins get spend controls in the Admin Console on day one: workspace-level defaults, group limits, individual overrides, and a process for reviewing requests for additional credits with user-submitted rationale. Vendors do not build credit-request workflows for features that are cheap and predictable.

!What OpenAI publishes about ChatGPT Work usage against what it does not: no per-task rate, no cost known before a run, no cost per firing of a recurring task\n\nThat reframes the planning question. It is not "can I automate this recurring job." It is "what does this job cost each time it runs, how do I find out, and who notices when a monitor set to check every morning quietly consumes a team's allowance."

Four other things worth knowing

Work is on every plan on desktop, including Free. The web and mobile rollout was staged (Pro, Enterprise and Edu first, then Plus and Business), but the desktop app carries Chat, Work and Codex "to users on every plan, including Free." If you concluded you did not have access, check the desktop app.

Auto-review sits between the agent and consequential actions. OpenAI describes using its most advanced models to review important actions involving connected tools and APIs before they happen, to prevent unauthorized sharing of sensitive information. That is the approval model doing real work, not a checkbox.

Plugins, not connectors. The vocabulary changed. Plugins connect Slack, Teams, Google Drive, SharePoint, email, calendars, CRMs and project trackers, in a unified directory. You can point at one explicitly by typing "@" and the app name. OpenAI does not publish a count, so treat any specific number you see as unsourced.

Atlas is being sunset. The standalone Atlas browser is going away, its capabilities folded into ChatGPT's built-in browser on desktop and an updated Chrome extension that runs in Chrome's sidebar. A product discontinuation, announced inside someone else's launch.

What to do with this

Pick the task you redo, not the task you dread. The seven above share a shape: repeated, structured, low-judgment. Dread usually signals the opposite.

Name the artifact. Columns, speaker notes, slide count, format. The output framing is the whole advantage, and it only works if you say what the output is.

Before you schedule anything, run it once manually and watch what it consumes. That is the only way to find out what a recurring version will cost, because the rate is not published.

If you administer a workspace, set spend controls before adoption, not after. They exist because this is the kind of feature that produces a surprising invoice.

Rewritten 8 September 2026 against OpenAI's launch announcement of 9 July 2026. Customer outcomes are OpenAI's own reporting of its customers and are labeled as such.