I lost a morning last week to a product name. A client asked whether her ops team should switch to “that new ChatGPT Work thing,” and I promised a one-paragraph answer by lunch. Six days later I’m writing a full post instead, because the honest answer to “what is ChatGPT Work” starts with a counter-question: which one? There are two different products called ChatGPT Work. They run in different places, they can touch different things, and OpenAI doesn’t go out of its way to tell you which one you’re using.
I have a stake in getting this right. My blog’s content pipeline, from source ingestion to published posts, runs through Claude’s Cowork with a cheap VPS doing the plumbing. When OpenAI ships something aimed at exactly this kind of delegated work, I want to know whether I picked the wrong horse. So I spent the week giving ChatGPT Work real chores instead of demo prompts. Here’s what I’d tell that client over coffee.
It’s two products wearing one name
Simon Willison has written the most careful public breakdown of this so far, and his naming is the one I’ll borrow: there’s Work Cloud and there’s Work Local. OpenAI announced ChatGPT Work on July 9th and has been reworking it almost weekly since, which is part of why it’s so hard to pin down. Willison calls it “an extraordinarily confusing and very powerful product,” and after a week with it I’d co-sign both halves of that sentence.
Work Cloud is what you get on chatgpt.com or in the mobile apps. Your task runs on OpenAI’s machines, in their sandbox, on their schedule. You hand over a job, close the laptop, and check back later.
Work Local lives inside the ChatGPT desktop app, which until recently was called Codex. It can read files and run programs directly on your computer. It feels like Codex re-skinned to be less intimidating for people who don’t write software, because that’s roughly what it is.
The split matters more than it sounds. Where an agent runs decides what it can touch and how bad the blast radius is when something goes wrong. A cloud agent can’t delete the folder with your invoices in it. A local one can.
Work Cloud is the half worth your attention
Inside chatgpt.com, Work shows up as a tab next to Chat. That’s a small interface decision with a loud message: OpenAI thinks chatting and working are different modes, and the second one is where the product is headed. Work also gets features that plain Chat doesn’t, so the tab isn’t cosmetic.
The pitch, once you strip the branding, is delegation. Chat is you plus the model, taking turns. Work is you handing over a task and leaving the room. In my testing, that difference was less about model quality and more about my own behavior. With Chat I hover. With Work I wrote a decent brief, went off to do something billable, and came back to review the output the way I’d review a contractor’s.
That review step is carrying a lot of weight in that sentence. Some jobs came back finished. Others came back confidently half-done, and the polish made the gaps harder to spot. If you’ve ever managed junior freelancers, the feeling will be familiar. The quality of your brief mattered more than anything else I changed. A vague ask produced a vague deliverable, every single time, and the failure showed up an hour later instead of in the next chat message, which makes vagueness more expensive than it used to be.
The pricing wall tells you the strategy
ChatGPT Work is for paid subscribers only, $20 a month and up. Free users don’t get it. Neither do the $8-a-month Go subscribers, and that’s the exclusion I find more interesting. OpenAI has drawn a line straight through its own paying customers: cheap conversation below it, agents above it.
I read that as a statement about where the money is. Conversation is a commodity now. Every lab gives away a decent chatbot. Delegated work is the thing you can invoice for, whether the customer is a freelancer like me or a government. The same week I was testing this, TechCrunch reported that versions of ChatGPT and Grok are joining Gemini on the Pentagon’s central AI portal. Different buyer, same motion: the chat is the demo, the work is the product.
ChatGPT Work vs Claude Cowork
This is the comparison I actually care about, since Cowork is what runs my own pipeline. The two products are converging on the same shape: give the model a computer, some tools, and permission to take many steps without checking in between each one.
The differences are mostly about where the computer sits. Cowork runs on my Mac and only sees folders I’ve explicitly handed it, plus a sandboxed Linux shell for the messy parts. Work Cloud runs on OpenAI’s hardware, which means less setup and less control, in both the good and the bad sense. Work Local is the closer cousin to Cowork, right down to the “this used to be a developer tool” ancestry.
After a week I don’t think either one is simply better. Work Cloud is easier to start with; there’s nothing to install and nothing it can break on your machine. Cowork fits me because my jobs need my files, my SSH keys, and my folder structure. The thing I’d push anyone to compare first isn’t the demo reel, it’s the permission model: what can the agent read, what can it run, and what does it show you afterwards. I wrote about how agent tool permissions go wrong a while back, and everything in that post applies double once the agent works unsupervised.
Should you pay for it?
If your day is mostly asking questions and drafting text, no. Chat already does that, and the free tier of any major assistant does that.
If your day contains chores with steps, the math changes. The test I give clients is simple: is there a task you do weekly that you could hand to a new assistant with one page of written instructions? If yes, an agent product will probably eat some of it, and $20 a month costs less than fifteen minutes of a freelancer’s time. My own pipeline is the proof I point at. It drafts, a quality gate checks the rules, and I read the output before it matters. That structure is what I build for clients when they want automation that doesn’t embarrass them.
What I wouldn’t do is buy the subscription first and go hunting for uses second. That’s how you end up with an expensive way to summarize emails you didn’t need to read.
What to try this week
Pick one recurring chore. Write the steps down as if you were briefing a new hire, including what “done” looks like. Give that brief to whichever agent product you already have access to, ChatGPT Work or Cowork or anything else, and grade the result like a manager, not like a fan.
If it fails, you’ve lost twenty minutes and learned exactly where your process was underspecified, which was worth knowing anyway. If it works, you’ve found your first delegated task, and the second one is usually easier to spot. Mine now drafts blog posts while I sleep, including chunks of this one, which felt strange to admit and then immediately normal.
