Somebody asked me last week which one they should be paying for. That question expired sometime this summer and most of us did not notice.
I run three right now. Claude Cowork, ChatGPT Work, and Grok Bot. One at a time, for one reason each, and then I looked up and I was on all three.
They have converged, and the convergence is the story. All three run skills, which is your written-down method in a form the thing can follow. All three take connectors, so they reach into systems you already pay for. All three drive a browser, so they can work a site that never built an integration for you.
Twelve months ago those were three different products. Now they are the table. Comparing feature lists is no longer how you choose, because the lists have stopped disagreeing.
So what do I use them for.
Most of my work still happens in Claude Cowork, because it works inside my own folder, on my own files, next to me while I do it.
I reach for ChatGPT more than I expected to. It is strong and noticeably faster, and on the days speed is the thing that matters, that is not a small difference.
Grok Bot gets the jobs that need an always-on connection to a browser. Something that keeps sitting there whether or not my laptop is open.
Three different jobs. None of them is the winner, and I am suspicious of anybody telling you one of them is.
Here is the part I want you to take with you, and it is not that the swap worked.
I pointed my own setup at a second one of these a while back. Same connection into my own systems, same tools, same written-down methods. I expected friction. There was almost none. So I did it again with the third, and all three took it.
For a while I assumed I had gotten lucky. Three vendors, three release schedules, no reason on earth for them to line up. Lucky is what it looks like from the outside.
I had not gotten lucky. The format those methods live in is an open standard. A skill is a folder with a markdown file in it. That is close to the entire specification, and three companies that compete directly with each other all agreed to read the same one.
That distinction matters more than the swap does. Goodwill expires. Architecture does not. If the portability were three vendors being generous, it would last exactly until one of them decided it was a differentiator worth taking back. Because it is a published format instead, taking it back means shipping something that no longer reads a file everyone else reads, and nobody wins that trade.
Which is why none of the parts that count live in a chat window. My records are in my own database. The connectors point at my own systems. The methods are files describing how I work. The assistant is the part doing the typing, and the part doing the typing is the replaceable part, on purpose, by design, in a format three companies published agreement to.
One more beat, because it cuts the other way. Before I tried the third one, I read three write-ups on it, and one said flatly that it does not take connectors. It does. I know because I plugged mine in and it worked in about a minute. Small story about one product, large story about how much of what you read on these tools is written by people who never ran them against their own account.
The play. Point your setup at a second one and see what survives.
You do not have to switch anything and you do not have to cancel anything. Take whatever you have already built, a Project with real instructions, a connector into your email or your drive, a procedure you have written down, and stand it up somewhere else for an afternoon.
Then look at what you had to rebuild. That list is the part you do not own.
If everything came across, your work lives in your own stuff and you are in a stronger position than you thought. If most of it had to be rebuilt by hand, that is not a failure either. It is just the map. Now you know which pieces to move out of the chat window first, and you found that out on a quiet afternoon instead of during the week a tool you depend on changes its pricing or its terms or shuts off the thing you were using.
I did not run that test to prove a point. I ran it by accident, one tool at a time. The reassuring part was not that it worked. It was finding out afterward that it was never going to fail, because the thing holding my work is a published format and not a favor.
Know somebody trying to pick one? Forward this. And hit reply: what would break first if you had to move? I read every one.
Written by Paul Cobb, real estate operator (Louisville / Kentuckiana), forVEX builder. I run the plays I write about. More depth: forvex.app/journal