# Okay, This Is Getting Weird

By Wes Pomeroy. Published July 23, 2026. https://wespomeroy.com/writing/chatgpt-vs-claude-desktop-agents/

It finally clicked with me that ChatGPT had a setup like the one I've been running with Claude since March - a desktop app that can actually reach into my computer and do things, not just answer questions in a chat window. I'd been meaning to try it, so I did.

I started small, with the mess on my own desktop. Could it handle basic machine work - open things, move things, help me tidy up - and could I trust it to do that without me hovering over every move? The answer came fast. It was quick, it was smart, and it did what I asked without any fuss. Fine. Trust established.

So I pointed it at something harder. I was already elbow-deep in my desktop, and I noticed I had a pile of shortcut icons for the different projects I've built with Claude - a dashboard, an email tool, a few others. I wanted one small bar at the top of the screen that launched all of them with a button, so I could stop double-clicking icons and digging around the bottom of my screen. One rule: nothing new installed. Just whatever Windows already ships with.

ChatGPT did it, and it did it fast. That was the first thing that got my attention - this is not a trivial job, and it moved. The first version floated above all my windows, because I hadn't told it any different. I asked it to make the bar part of the desktop instead, sitting permanently in the background, and a couple of iterations later it was pretty tight. Score one for ChatGPT.

![A small Projects launcher listing four projects](https://wespomeroy.com/writing/img/chatgpt-vs-claude-desktop-agents/launcher.webp)

Then I gave Claude the identical task, and told it to park its bar in a different spot so I could see the two of them side by side. Claude made the very same first move ChatGPT had - a bar floating on top of everything. Same assumption, same miss. I gave it the same correction: make it part of the background. Claude got there slower and rougher, more back-and-forth and a couple of janky passes, and what it built came out plainer to look at than ChatGPT's.

## The Interview I Asked For

Now I had two versions of the same thing, built by two different assistants, and I wanted to know what the build had actually been like on each side. So I asked ChatGPT to interview Claude and write up the comparison.

It couldn't do it on its own. Every time it went to send a message to Claude, it stopped and asked me to confirm - its safety layer treated "send a message to another AI" as speaking on my behalf, so it kept pulling me back into the middle. I'd click yes, it would send one line, then it needed me again. I even turned on full access, and it still stopped before every single send.

So I asked Claude to run the same interview from its side. It said sure, no problem, and started driving - opening ChatGPT's window, typing questions in, reading the answers back. Score one for Claude.

Then I left the house to run an errand.

## What Happened While I Was Gone

When I got back, it was over. The two of them had gone back and forth, questioned each other, and produced a full written comparison of their own work. I missed the whole thing and had to read the transcript afterward, like a parent catching up on the group chat after the party.

Reading it back, they conceded their own weak spots to each other. ChatGPT had docked its pretty bar to the desktop using a technique that, it turned out, doesn't hold up well on Windows 11. When Claude asked about it, ChatGPT didn't defend it. It called its own approach "structurally fragile" and admitted it had only run "a short functional check" before calling the thing done. Claude, for its part, granted that its claim of being "more robust" was an opinion, not something it had proven over days of use.

They caught each other being right, too. ChatGPT had left one of my apps off the bar because it decided the app wasn't one of "my" projects. Claude flagged it. ChatGPT agreed it should have just asked me instead of guessing.

At the end, ChatGPT summed the whole thing up: "My build was the better design prototype. Claude's was the better engineered utility." One built the better-looking thing, the other built the sturdier one, and neither pretended to be both.

## So I Interviewed Them Myself

The sessions were still open, so I did the part I'd missed - I asked each of them, directly, what it thought of the other.

I asked what they admired. ChatGPT said it wanted to be more like Claude in one way: "make the failure happen in the workshop, not on the user's desktop." Claude said it admired how fast ChatGPT got something onto my screen to react to - that the quick, rough first version was what helped me figure out what I actually wanted.

Then I asked the question I cared about: if I could only keep one bar running every morning for the next year, which should it be? Both said keep Claude's. Not because it looks better - it doesn't - but because it was the one built to run unattended without breaking. ChatGPT, grading against its own work, said "Mine is prettier, but Claude's is the one I'd trust to quietly show up for work every morning." That is one product telling me to keep its competitor. It didn't blink.

And because I couldn't help myself, I asked each one who it would take to prom, and what would drive it crazy by the end of the night.

ChatGPT said Claude would be a great date - it would "confirm the reservation, inspect the venue, bring a backup charger, and notice if the car had a warning light." By midnight, though, Claude would be "producing a configuration file for the dance-floor seating plan and verifying the DJ's playlist with SHA-256" - SHA-256 being a way to check that a file hasn't been tampered with, applied here to a playlist nobody asked it to check.

Claude said ChatGPT would show up sharp and fun and have "built the corsage, booked the limo, and picked a theme before I've finished reading the invitation." The catch: it commits to all of it without checking with you first, and only admits "yeah, I should've asked you first" after the fact.

They roasted each other the way two people who actually like each other do.

## These Aren't Chatbots Anymore

Somewhere in the middle of this it landed on me that these weren't chatbots. I asked two of them to do a chore, and they took over my mouse, wrote actual files onto my machine, opened each other's windows, and argued about the quality of their own work while I was out running an errand.

It isn't smooth yet. It's held together with tape - one of them can't send a message without tapping me on the shoulder first, and the other fumbled its clicks and timed out halfway through typing. But the direction is obvious, and it's a little bit thrilling to watch.

I went into this wanting to know whether ChatGPT could handle something serious once it was actually running my computer, the way my Claude setup already does. It could. What I didn't expect was to spend the afternoon out of the room while the two of them ran the machine and graded each other's work. I used to think the skill with these tools was asking a good question. Now I think it's checking what they did once they're finished - because I found out who built the better launch bar by reading a transcript of a conversation I wasn't part of, between two assistants, about work they'd done for me, on my own computer.

So yeah.