Around Chinese New Year this year, OpenClaw suddenly took off. I gave Agent‑building a shot myself, and spent the better part of a day getting tangled up in all sorts of small issues. After that I switched to OpenAI's Codex desktop app, which was a lot more stable — it's an official app, not a bug pit. Already in a different league from OpenClaw.
We'd actually tried Claude earlier than that, but at the time it only worked through code mode — you had to ask questions inside the Mac Terminal. The barrier and ongoing friction were both high enough that I didn't stick with it. About a month later the Claude web version came out, and that's when it really came back onto my desktop.
Codex is a developer, Claude is a boss
After using both for a while, the difference between Codex and Claude turned out to be huge:
- Codex feels more like a developer. You have to spell the requirement out clearly. It can take you from 0 to 90% of the work, and that part is genuinely strong — but the last 10% you have to grind through over and over, because it has a hard time catching what you actually meant.
- Claude has a "thinking layer." It feels more like the boss of a web‑dev shop, or a senior engineer — closer to how a person actually reasons. That positioning drops your communication cost a notch, and the throughput ends up much higher than Codex.
As a production tool, once the efficiency gap opens up, the slower one quietly gets shelved. I'm on the 200 € / month Claude plan, and there are still moments when it feels a little tight.
My workflow: Claude dispatches, Codex grinds
My setup right now is fixed:
- Claude breaks tasks down, plans, makes judgement calls, and writes anything that needs thought
- Codex handles repetitive mechanical work — batch edits, batch runs
The result: I burn less than 5% of my Codex quota, while Claude runs at close to 100% almost all the time. In other words, I've gone from playing around to heavy dependency.
Claude and ChatGPT are different animals
People tend to lump Claude and ChatGPT together. They're really not the same thing.
ChatGPT is a Q&A tool — you ask, it answers. What it gives you is ideas. Claude is a production tool — you give it a task, and it does the work — it hands you a result.
That's a difference in kind, not in degree.
More than 12 hours a day
Since Claude moved into my daily workflow, I use it from the moment I wake up until I go to sleep. Excluding sleep and meals, I'm on it more than 12 hours a day — heavy user, no qualifier.
The effect on my routine is real. I used to keep a "normal" workday; now everything I do orbits around Claude.
I'd bet more and more people end up like me: using Claude the way we scroll a phone, to the point where not using it feels uncomfortable. The dependency these agents create is only going to get stronger.
On the valuation
Plenty of people say Anthropic is overvalued, that it's part of the AI bubble.
From my own usage, it's not a bubble. The reason is simple:
- The more you use it, the harder it is to walk away
- And this kind of user is willing to keep paying for it
The two days with Fable
One more thing made me even more confident in Anthropic's upside.
They released a model called Fable. I only got to use it for two days before it was pulled, without even really getting a feel for it. But even in those two days, the level was clear — it was among the strongest models on the market at that moment.
Putting a model of that caliber out in the open and then quietly pulling it back tells me, in reverse: the real hand Anthropic is holding is probably much bigger than what's visible from the outside.
That's exactly why I say OpenAI is in a fragile spot. Not because ChatGPT collapsed overnight, but because Anthropic can drop Fable, walk it back, and do it that calmly — what else is sitting in the other hand?
At this pace, I think it's plausible that Anthropic ends up overtaking OpenAI. Because ChatGPT and Codex are already pushed aside in my own setup — a few more days like this and, if I still haven't reopened them, I'll cancel both subscriptions.