ChatGPT Work is the agent mode inside the new ChatGPT desktop app. Instead of answering questions the way the chat you know does, you describe an outcome, it reads the files and apps you have connected, shows you a plan, and then does the work: documents, spreadsheets, reports, invoices, websites, entire multi-step projects. It arrived on July 9, 2026, when OpenAI merged the Codex app into ChatGPT and renamed the old desktop app ChatGPT Classic.
I recorded a full walkthrough the week it launched, on my own machine, with my own business connected:
Why This Announcement Is Different
The numbers don't lie. Almost one billion people use ChatGPT. Only about five million people per week were using Codex, OpenAI's agent. Five million sounds big until you put it next to a billion.
Now all of those ChatGPT users have the agent in front of their face. They are going to get a taste of what it feels like for an agent to take work to the finish line, instead of them taking the idea and finishing the work themselves. And it only takes a few of your competitors getting that taste.
Here is the theory I keep coming back to. The businesses that got on the internet early, that planted their roots and built a presence while their competitors debated whether it mattered, held an advantage for a decade. The same thing is happening with agents. The businesses that restructure their operations to be supported by agents earliest and best will hold an advantage that will not even be fair.
You could not do this from the old ChatGPT desktop app. You could get copy for a website, but the app could not build the website, deploy it, match it to your branding, and keep it updated as your business changed. That was the dinosaur age of using AI, and it ended on July 9.
What ChatGPT Work Actually Does
The new app has three modes at the top: Chat, Work, and Codex. Chat is the ChatGPT you know. Work and Codex are the agent, and honestly, on this interface there is very little difference between the two beyond the starter prompts. One app, one agent, two doors.
What changes for you as an owner is the shape of the output. A chat gives you an answer. The agent gives you a finished thing:
| You say | It delivers |
|---|---|
| "Send this client an invoice for the deal we closed today" | The invoice created and sent through your connected payment tool, plus the email |
| "Restructure my Google Drive so humans and agents can both find things" | The reorganized Drive, with naming standards documented for the next agent |
| "Are we undercharging based on what we deliver clients?" | An analysis built from your contracts, emails, and call transcripts, with numbers |
| "Find the proposal we sent that logistics company in 2018 and build a new one from its structure" | The old proposal located, the new one drafted from its framework |
That last one is a real client story. He needed a proposal from eight years ago as the framework for a new deal in a similar industry. The agent found it in minutes and worked from it. No human dug through folders.
The quality of all of this depends on one thing: context. The agent can only be as useful as what you let it see. Which brings us to the map.
The Complete ChatGPT Work Guide for Business Owners
I have broken the full playbook into focused guides, each one answering the question owners actually ask me:
The One Principle Under All of It
Most people go to the agent with a solution. I rarely do. I give it the vision or the problem, admit I might be wrong, and let it propose the path. If you always arrive with exactly the solution you want, you inherit the ceiling of your own ideas, and the honest answer is that you are probably wrong about the best way to do it.
Here is a first prompt in that spirit, ready to paste once you have connected your files:
Paste into Work mode
Please suggest a better way to structure the following for human and agent work: local folders and files, Google Drive, and email organization. I am going to be using agents a lot more, from more than one AI tool, so the structure needs naming standards any future agent can follow. Propose the structure first, then test it: spawn sub-agents and see whether they can find things you make deliberately tricky to find, and show me the results before touching anything real.
That prompt made the agent build an adversarial retrieval test on my own files: a misleading file labeled final, an old asset labeled current, two similarly named events. Every sub-agent passed by following the rules instead of trusting file names. That is the difference between organizing for humans and organizing for agents, and it is the foundation everything else in this cluster builds on.
Common Questions
What is ChatGPT Work in one sentence? The agent mode inside the new ChatGPT app: you describe the outcome, it does the work across your files and connected apps.
Is it the same as Codex? Same app, same agent, two mode labels. The full answer is here.
Do I need to know how to code? No. Permissions and plugins are toggles and sign-ins. The setup tutorial takes about fifteen minutes.
What does it cost? Most owners start on the $100 per month plan and it carries the month. Heavy operators move to $200. The pricing guide covers credits, resets, and the overage trap.
What should my team use? Owner first, team second, with rules. The team rollout guide explains why, including the stories of what happens without rules.
Bottom Line
A billion people just got an agent put in front of them. Most will poke at it. A few of your competitors will restructure around it.
Be the one whose roots are already down when they start.