I've been dual-fisting Codex and Claude Code while building SignUppity, and I want to talk about what that's actually like.
First, the setup: I moved away from the agent-in-the-IDE workflow and started working from the desktop applications. Both of them. Simultaneously. This felt smart for about forty-five minutes, then I spent the next hour wanting to scream what is actually happening in the terminal right now at my laptop screen. The harness isn't everything, but it's a lot. When you give up the IDE integration, you give up your peripheral vision. Things are happening. You can see that things are happening. You just have no idea what.
I lost count of how many times I hit that wall. It was more than ten.
The Platform Arms Race Is Real and Worth Watching
Here's what I didn't expect: being a spectator to a feature parity war in real time is genuinely interesting. Claude and Codex are clearly pushing each other. You can feel it. OpenAI wants to catch up, and the pace of changes from both sides over the past few months has been relentless. If you're building on top of these tools, the ground is moving underneath you constantly, and I honestly don't know the best way for us little folks to ride that wave. I'm mostly just trying not to fall off.
Model-to-model, they've been more comparable than I expected. The bigger difference has been the economics. My five-hour credit window runs out faster with Claude, Opus 4.6 through 4.8 are absolute beasts with credit appetites to match. GPT-5.5 only occasionally feels outrageous by comparison.
Codex Had Two Genuinely Good Moments This Week
The first one was almost fun to watch. I needed a favicon. Codex took a PNG, stripped the background, converted it to an SVG, and handed it back like it had a BFA in graphic design and had been waiting for me to give it something interesting to do. Clean. No drama.
The second one is smaller, but I keep thinking about it.
I hit my usage limit mid-session. With Claude Code, that's it. It cuts off mid-token, mid-thought, mid-whatever-it-was-doing. You find out you are done when it stops. With Codex, when you hit the limit, it finishes the thought. It had one more thing to complete, and it completed it while telling me I was out of messages. Then it stopped.
That's a tiny thing. It's not a feature in any list. But the difference between a tool that abandons you mid-sentence and one that finishes the sentence before it leaves is the difference between a collaborator and a vending machine that ran out of quarters. Anthropic, I say this with love: please fix this.
So Who's Winning?
Neither. Both. Us? Them? The question's probably wrong.
What I'm actually learning is that these tools have different personalities even when their outputs are comparable. The personality matters more than I thought it would when you're spending even an hour a session with them. The terminal visibility problem is real and I haven't solved it yet. The credit economics are real and I'm learning to plan around them.
SignUppity is moving. The harness situation will get sorted. And if Codex keeps doing favicons like that, I might owe it a portfolio credit.
What's your experience been, dual-wielding these tools? I'm genuinely curious whether I'm the only one who finds the desktop workflow this disorienting.