I have been using Claude Desktop and Codex Desktop side by side. On a feature checklist, Codex Desktop clearly comes out ahead. In day-to-day use, though, I still find Claude Desktop more comfortable to work in.
At first I couldn’t explain why, so I want to break it down: where Codex is stronger, and where exactly the discomfort comes from.
Credit where it’s due
There are a few things I have to admit.
First, Codex Desktop currently has many more features than Claude Desktop.
Second, computer use in Codex works much better than it does in Claude.
Third, from my actual programming work, Astra 6 is stronger at frontend development than Opus 5.
On those three points alone, Codex Desktop should be the better choice. The problem lies outside the checklist.
The interface feels heavy
My first impression of the Codex Desktop UI is that it looks clumsy and heavy. It resembles an IDE, which isn’t necessarily a bad thing for a developer tool, but it doesn’t feel light to work in.
The interface of an AI desktop app is an environment I spend long hours in. However capable the model is, a heavy environment affects whether I want to stay and work there.
The input box doesn’t help me write clearly
The strangest part for me is the input box in Codex Desktop.
In Claude Desktop, I naturally write organized input: I lay out the context, the goal, and the requirements before handing the task to the model. In Codex Desktop, the input box doesn’t guide me toward writing that way.
It’s a hard feeling to describe. Both are text boxes and both accept typing, yet one makes me put care into how I express a task and the other doesn’t.
I think this deserves its own section, because giving an agent well-organized input is part of the work itself. An input box isn’t just a place to put text; it also shapes how I think through and describe a task.
“Slow” is a feeling
Codex Desktop also feels slow to me.
Looking at the models alone, though, both should handle problems at roughly the same speed, and Astra may even be a little faster. This is my judgment, not the result of a timed test.
In other words, the slowness I feel probably doesn’t come from the model. It comes from the product as a whole: a heavy interface and an awkward input experience make the whole process feel slower than it really is.
Where I’ve landed
This comparison made one thing clearer to me: how good an AI tool is to use depends on more than model capability and the number of features.
Codex Desktop has clear advantages in features, computer use, and frontend ability. Claude Desktop, with its interface and input experience, is the one that makes me want to sit down and work in an organized way. For a tool I use every day, that sense of comfort matters just as much.