Follow ZDNET: Add us as a preferred
ZDNET’s key takeaways
- ChatGPT Work can browse, manage files, and complete tasks.
- Free users can try it, but paid plans offer more capacity.
- Start small, limit permissions, and double-check its work.
ChatGPT answers questions. ChatGPT Work does work. That’s because ChatGPT Work is agentic, which is a fancy way of describing how some AIs can go off and accomplish a series of tasks.
Because ChatGPT has become more capable, the differences between ChatGPT and ChatGPT Work can be subtle. For example, if you ask ChatGPT to plan a travel itinerary, it will search the web and gather together a list of great destinations. That’s the AI doing something.
Also: ‘I can’t stop’: 80% of developers find AI coding more addictive than helpful
It gets even more blurry with ChatGPT’s Deep Research mode. Deep Research is essentially a special-purpose agentic web researcher. Deep Research can dig around the web and build a much deeper analysis than ChatGPT.
ChatGPT Work can do all that, and more. It can open your own Web browser, or its own internal one. It can move around the web on screen (rather than or in addition to retrieving a web page’s HTML data). It can create and move files. It can process through a long sequence of steps.
This approach is very much like what an AI agentic programming tool like Claude Code or OpenAI‘s Codex can do. In fact, agentic work tools like Claude Cowork and ChatGPT Work evolved directly out of agentic coding tools that some users decided to use for non-coding tasks.
Also: This AI expert says the job apocalypse isn’t coming, even if you’re a coder – here’s why
Following a surprising usage surge in non-programmer use of those programmer tools, both Anthropic with Claude and OpenAI with ChatGPT launched agentic work products. Today, we’ll investigate starting with ChatGPT Work.
What you need for ChatGPT Work
ChatGPT Work is available for all ChatGPT users, even those on the free tier. There are limits, though. Agentic operations take a lot of AI juice (typically measured in tokens). Free users are given access to only a small amount of capacity. As you pay more, you get more usage capacity.
I have been using the $20/mo ChatGPT Plus, which has proven to be pretty capable. When I use Codex for coding, it tends to slam shut after about five hours of intense work. Still, that’s five hours of intense work for $20. That usage limit resets after a while. I’ve found I can get about 20 or so hours of agentic coding done each month on that tier.
ChatGPT Work uses a similar amount of capacity. You can probably do a fairly large project with Work every few days and not feel all that limited on the $20/mo plan. Unfortunately, OpenAI (and this is true of all the AI companies) doesn’t have a single metric that you can measure by. Each request requires a different mix of re
But I’d think of it this way: free gets you a taste, $20/mo gets you fairly regular moderate usage, and the higher tiers let you binge.
Also: 75% of developers I surveyed prefer Claude Code – here’s why they choose it over Codex
Beyond your account, you’ll need to either sign in to the ChatGPT web page or download the desktop app. For any tasks you want to run on your computer, you’ll need the desktop app.
Note that up until mid-July, the ChatGPT desktop app was a chat app. Then, in July, OpenAI combined ChatGPT, Work, and Codex into a “super app,” which had the unfortunate side-effect of killing some of the ChatGPT app’s best features.
So now, if you want to use ChatGPT Work on the desktop, download the ChatGPT desktop app.
Also: I loved ChatGPT Desktop until OpenAI gutted it to make room for Codex and Work
You can ignore my grumbling about the removed features. OpenAI still hasn’t bothered to put them back in. But I’m not bitter. It’s been almost two months. I’m not holding a grudge. Not at all.
How to use ChatGPT Work
Once you launch ChatGPT using either the desktop app or in your browser, you’ll see two buttons above the chat message bar. One is Chat, and one is Work.
Clicking the buttons (above) changes the mode. On the top, we’re in Chat mode. On the bottom, we’re in Work mode. Notice also the option to choose the folder you’re working in on the bottom. This approach means you can restrict a complex work project to a single folder or hierarchy.
Also: A low-tech solution from the past may be your best defense against AI deepfakes
Now, all you do is prompt the AI just like you always have. The difference is that Work can do a lot more. Last month, for example, I had ChatGPT Work dig through a large collection of PDF files, rename them, and organize them in topics. You can read about how I approached this project here.
One thing that did make me nervous is that Work hasn’t been asking for permissions as often as I’d like. I’ve generally learned to tune my requests so it doesn’t have a reason to run amok, but I would prefer to be asked for an okay more often.
An example research project
Before I set you loose on ChatGPT Work, I’d like to show you a research project I gave the tool. I’m planning an article comparing vertical mice, and I want to test a bunch of them to see which works best.
But I don’t just want vertical mice. I want wireless vertical mice that can switch between two to three devices (three preferred). I started digging through Amazon and realized I’d spend an hour or more trawling specs because the product headline doesn’t always indicate how many devices a mouse supports.
Also: You can earn your Google Al Professional Certificate for free – and I highly recommend it
Then I remembered that I have an agentic AI at my disposal. It was time to delegate.
I gave ChatGPT Work one prompt that consisted of the following components:
I want you to do some Amazon research for me. I want you to find wireless vertical mice.
That was my basic overview. You’re going to research mice on Amazon. Then I discussed the key restriction that would narrow the search and save me time:
They must be able to switch between at least two Bluetooth wireless devices (but I prefer three). Ignore vertical mice that aren’t wireless and ignore vertical mice that only support one computer.
Finally, I gave it an assignment only ChatGPT Work (and not ChatGPT or ChatGPT Deep Research) could do. I told it to check each candidate for Prime shipping and fairly immediate availability, and then open each one in its own tab. I wanted to quickly toggle between tabs, close ones I was uninterested in, and then narrow my evaluation scope:
When you find them, open each candidate in its own tab. Only choose mice that are Prime and will ship in less than a week.
As you can see in the image below, that process took ChatGPT a little over 10 minutes, during which I went and got myself a nice cup of coffee and played with the dog:
Ten quick tips
Here are some tips I recommend, based on my experience:
- When possible, feed the AI small assignments one at a time.
- For longer projects, ask the AI for a plan before it executes anything.
- Check your usage availability in Settings > Usage. If you’re about to give the AI a big project and your usage availability is low, it will stop mid-run. That sucks.
- Don’t give the AI any passwords. Ever.
- If you’re going to have the AI do anything destructive (like reorganize a folder), make very sure you have a backup or five. And make sure the backups are well out of the AI’s reach.
- Don’t believe anything the AI tells you. Double-check everything.
- Don’t be upset if the AI is stupid. Expect it to be. The AI can also be very helpful, but you have to pace and lead it with care and attention.
- If you’re about to start a long, tedious task, ask yourself if the AI can step in instead. Delegation is a good skill.
- If you decide to use the Gmail connector to send email messages, have the AI create drafts. Review each draft before pressing send.
- When you set up a connection, you need to grant permissions. Give the fewest permissions you can. It’s fairly safe to allow reads. But anything that involves writing, deleting, sending, or otherwise taking an action can backfire.
Well, there you go. I’ve found that I’m using ChatGPT Work and Claude Cowork more and more. I switch between the two mostly so I can share my experiences with you. Frankly, my experience has been that both are pretty much on par.
And with that, let’s call it a day. Have you used an agentic AI coworker like ChatGPT Work? Let us know in the comments below.
You can follow my day-to-day project updates on social media. Be sure to subscribe to my weekly update newsletter, and follow me on Twitter/X at @DavidGewirtz, on Facebook at Facebook.com/DavidGewirtz, on Instagram at Instagram.com/DavidGewirtz, on Bluesky at @DavidGewirtz.com, and on YouTube at YouTube.com/DavidGewirtzTV.
Artificial Intelligence
How to avoid Claude watermarking your content
Taught by AI pioneers, Stanford’s free online course takes you far beyond ChatGPT
75% of developers I surveyed prefer Claude Code – here’s why they choose it over Codex
How to keep your conversations with ChatGPT, Gemini, Copilot or Claude as private as possible
- How to avoid Claude watermarking your content
- Taught by AI pioneers, Stanford’s free online course takes you far beyond ChatGPT
- 75% of developers I surveyed prefer Claude Code – here’s why they choose it over Codex
- How to keep your conversations with ChatGPT, Gemini, Copilot or Claude as private as possible
- My 3 favorite AI tools for voice dictation while vibe coding – and one is free
