Close the Lid

01The sentence that sells it

Every vendor sells cloud scheduling with the same line. Anthropic’s Cowork help page says scheduled tasks “run on their cadence even when your computer is asleep or the Claude Desktop app is closed.” Its routines page opens by promising they “keep working when your laptop is closed.” OpenAI’s scheduled-tasks page makes the same point without the slogan: the task runs on its servers, and you get a notification when it is done. Every write-up of the feature repeats that idea as the obvious advantage, and I understand why: it is the sentence that makes the feature feel like magic.

I read it as the obvious liability. The property being advertised is that the work continues without you. The property I actually want, once I have more than a handful of these things, is that the work stops without me having to ask anyone. A lid does that. A settings page, eventually, after a login and a scroll and a confirmation, also does that. Those are not the same product.

The vendor wrote “closing the laptop lid still puts it to sleep” as a caveat. It is the best feature on the page.

02What “the limit” actually is

Every paid Claude plan meters you two ways: a session limit on a rolling five-hour window, and a weekly limit on top of it. The pricing page says web, desktop, mobile and Claude Code all draw from the same pool, and the routines page says routines “draw down subscription usage the same way interactive sessions do.” So the chat you are having, the code you are writing, and every scheduled task you have ever created spend from the same account, and the scheduled ones spend while you are not looking.

When a cloud task arrives at that limit there are exactly two outcomes, and the routines documentation states them without decoration. If usage credits are turned on, you “can keep running routines on metered overage.” If they are off, “additional runs are rejected until the window resets.” The first outcome is a bill at API rates. The second is a lottery: which of your tasks got through before the meter tripped is decided by the clock, not by you. Neither outcome is under your control at the moment it matters, and the moment it matters is usually 2 a.m.

The desktop offers a third outcome, and it is documented just as plainly, though it is written as a warning. From the Claude Code Desktop page: “Tasks only run while the desktop app is running and your computer is awake. If your computer sleeps through a scheduled time, the run is skipped.” And, two sentences later, the line I quoted above: “Closing the laptop lid still puts it to sleep.” Cowork says the same thing in its own interface. Point a task at a folder on your Mac and the app labels it “Only on this computer,” with one line underneath: “Stops when the app closes or this computer sleeps.” Skipped is not rejected. Skipped costs nothing, queues nothing, and requires no one’s permission.

03The brake is bounded, too

The objection I expected, and the one I had myself, is that a paused schedule is a debt. Close the lid for a weekend and Monday morning becomes a stampede of every run you skipped. That is not how the desktop scheduler works. When the app starts or the machine wakes, it checks the last seven days, starts exactly one catch-up run for the most recently missed time, and discards everything older. The documentation’s own example: “A daily task that missed six days runs once on wake.”

So the lid is not a pause that accrues. It is a pause that forgets, on purpose, and forgetting is what makes it safe at scale. Compare the cloud version of the same weekend. Runs rejected at the limit are not skipped; they resume at full cadence the moment the window resets, in the same pool your own Monday morning work is drawing from. You did not save anything by hitting the limit. You deferred it into the next window, and the next window is one you wanted for yourself.

Figure 01 Four places a scheduled task can live, and what each one lets you do when the meter trips

Property	Claude desktop (local task)	Claude cloud (routine / remote Cowork task)	ChatGPT desktop app (local task)	ChatGPT cloud (Scheduled)
Runs on	Your machine	Anthropic-managed cloud	Your machine	OpenAI’s servers
Needs the machine awake	Yes. Lid closed means asleep means skipped	No. That is the selling point	Yes. “Keep the computer on and the app running”	No. Runs on OpenAI’s servers and notifies you
Fastest cadence	Every minute	Every hour; faster cron expressions are rejected	Minute-based intervals inside a chat; custom RRULE otherwise	Hourly on paid plans; once a day on Free
Access to a folder on your computer	Yes, and a folder is required	No. Cowork’s remote tasks “can’t be tied to a folder on your computer”	Yes, the project directory or a worktree	No. Uploaded context and connected tools only
Permission prompts	Configurable; serious users turn on “Skip all approvals” and it never pauses	Routines: “no permission-mode picker and no approval prompts during a run.” Remote Cowork tasks carry an approval-mode setting	Your default sandbox settings	Some actions pause for approval
How many can you hold	No cap documented	A daily cap on runs per account	No cap documented	15 active tasks on Pro; 5 on Plus; 3 on Free
Past the plan limit	Whatever you decide, because you are there	Metered overage if credits are on; rejected if off	Whatever you decide, because you are there	Not spelled out; credits are documented for Codex, Work and Excel
The brake	Close the lid	Log in, find settings, turn credits off or pause tasks one by one	Close the lid	Pause tasks one by one

04The arithmetic at scale

Here is where the mainstream advice and I part company for good. With five tasks, none of this matters. Five daily briefings will never push a Max plan past its limit, and if they did the overage would be pocket change. The advice to “just use the cloud” is correct for the person it is usually given to.

It stops being correct somewhere in the low hundreds. A scheduled task is not a message; it is a session. It reads its connectors, searches, drafts, revises, and writes a report, and a session that touches a large context can cost real money at API rates. Take a deliberately ordinary run on Opus 5: 150,000 input tokens at $5 per million and 5,000 output tokens at $25 per million comes to roughly ninety cents. And that is at Opus rates. We now live in a Fable and Astra world: Claude Fable 5.1 and GPT-6 Astra, the two models everyone will actually want running their tasks, are both priced at $10 per million in and $50 per million out, so double the figure if you pick either. Now put a number of tasks next to it.

Figure 02 Monthly exposure if a day’s runs land past your plan limit, at about $0.90 per run

20 runs a day $540

100 runs a day $2,700

500 runs a day $13,500

2,000 runs a day $54,000

• Past the point where a plan can absorb it
• Mostly absorbed by a Max plan

Read the middle of that chart, not the ends. At a hundred runs a day, which is one person with a serious habit and not a company, the exposure is a few thousand dollars a month on top of the plan. That is the number the “runs while your laptop is closed” sentence is quietly describing. It is not a hypothetical either; it is precisely the scenario the routines page anticipates when it explains how to turn credits on so runs keep going past the limit.

The lid moves the cutoff to wherever you want it. Your plan’s included usage still absorbs the first slice, the same as it does in the cloud. The difference is what happens to the second slice. In the cloud it becomes either a charge or a rejection, chosen by the meter. On the desktop it becomes nothing, chosen by you, with a gesture that takes less than a second and needs no password.

05The objections

1. “My laptop sleeps on its own.”

   Then tell it not to. The desktop app has a “Keep computer awake” setting for exactly this. With it on, sleep stops being an accident and becomes a decision, and the decision is the lid. That is the whole design: the machine runs until you physically say otherwise.

2. “I need it to run at 9:00 sharp.”

   Some tasks do, and the vendor’s own advice covers them: put the guardrail in the prompt. “If it’s after 5pm, skip the review and just post a summary of what was missed.” Most scheduled work is “sometime today,” and a task that ran at 7:40 because you opened the lid at 7:40 has lost nothing.

3. “The cloud is more reliable.”

   Reliability you cannot switch off is a liability once it costs money. And read the fine print on that green checkmark: a green cloud run means “the session started and exited without an infrastructure error. It does not mean the task in your prompt succeeded.” A skipped desktop run is at least logged with its reason.

4. “What about ChatGPT and Astra?”

   Same argument, bigger numbers. GPT-6 Astra shipped this week, “included within the existing subscription allowances,” with credits for sale “for additional usage,” at $10 in and $50 out on the API. The ChatGPT desktop app, which replaced the Atlas browser this summer, has the same local mode with the same rule: keep the computer on and the app running. Cloud scheduled tasks cap at 15 active on Pro. If you have dozens, the desktop is the only place they can live anyway.

06How to do it

The good news is that none of this requires a workaround. Both vendors ship the local mode; they just do not lead with it. In Cowork the rule is written into the manual-setup instructions: “If a scheduled task requires local files or apps, it will only run locally.” So give every task a folder. On my Mac, a task pointed at a folder shows the “Only on this computer” label, and a task that only exists on this computer obeys the lid. In Claude Code Desktop, the New Routine dialog asks you Cloud or Local; pick Local. In the ChatGPT desktop app, pick a project directory.

• Give every Cowork task a folder, which forces it local
• In Claude Code Desktop choose Local, not Cloud
• In the ChatGPT desktop app, pick a project directory
• Turn on Keep computer awake
• Turn on Skip all approvals so tasks actually run unattended
• Leave usage credits off, or capped low
• When the limit notification lands, close the lid

A word for readers who are new to this, because the setting matters here. Out of the box, Cowork pauses to ask before it does anything consequential, and a paused task is a task that is not spending. That is not how people with dozens of tasks run it. The first thing a serious user does is turn on Skip all approvals, and the app is candid about what that means. The banner reads: “Skip all approvals is on. Claude never pauses, even for unsafe actions. This includes using your connectors. You can turn off individual connectors in the + menu.” Nobody running three hundred tasks is going to sit and click “allow” three hundred times a day, so that banner is on for every sophisticated user I know, and it should be. Approvals are a brake for one task at a time. They are useless as a brake for a fleet.

Which means the human is already out of the loop on both sides. A cloud routine runs with “no permission-mode picker and no approval prompts,” using every connector you attached unless you remembered to remove it. A desktop task with approvals skipped runs exactly the same way. The only difference left between the two is whether there is a hinge between the task and your wallet, and only one of them has one. Once you have turned approvals off, and you will, the lid is not one of your brakes. It is the brake.

07The ledger

Forecasts should be checkable, so here are the claims with their failure conditions attached.

08What I left out

Cowork’s own help page does not say what a remote scheduled task does at the plan limit. The routines page does, and I have applied that rule to Cowork by inference because both run on the same account and the same pool. If Cowork behaves differently, that is the first thing I would want corrected.

On the OpenAI side, credits are documented for Codex, ChatGPT Work and ChatGPT for Excel. Whether an ordinary cloud scheduled task can draw on them is not spelled out. Under a fixed plan the cloud failure mode there may be a silent pause rather than a bill, which is cheaper, but no more under your control. I also left out Team and Enterprise admin spend caps, which exist and which change the arithmetic for organizations; this piece is about an individual on Pro or Max. And I left out the electricity and the wear on a laptop kept awake all day, which are real and which round to zero against API rates.

The industry’s line is that the cloud lets the work continue without you. Fine. What I am paying for is the ability to make it stop without asking anyone, at the moment I decide, with my hand. That ability is a hinge.

The industry sells “runs while your laptop is closed” as the feature. I think the feature is that it doesn’t.