For months people have been telling me Codex is better than Claude Code.
I ignored all of it. I have built my whole business inside Claude Code (every pipeline, every system) and I did not fancy starting again.
Then GPT-6 Astra launched and I (finally) designed to make the switch.
The one thing keeping most people where they are is setup debt.
You have spent months teaching a tool how you work. Your rules, your instructions file, the way you brief it.
Moving means starting again on all of it and going back to generic output. That sunk cost is real, and it is why people stay.
Good news: OpenAI just made it a four-minute job!
Codex finds the agents already on your machine and offers to bring them across. Your setup, your skills, your last 30 days of chats.
And if you have never used one of these at all, this is the easiest moment there has ever been to start. It’s ChatGPT that can see your folders and vibe code stuff.
Instead of pasting a document into a chat window, you point it at the folder the document lives in. That’s it.
Step 1. Import it in the app.
Download Codex and open it. There is nothing to install in a terminal and no command to type (which is why most people avoid Claude Code).
The route is Settings, then Import, then Claude Code. It scans your machine on the way in and shows you what it found. Mine offered all three.
Nothing is moved and nothing is deleted. You get a copy, and Claude Code carries on exactly as it was. Then it asks what to bring over. Tick everything.
Your rules file gets renamed on the way in, from CLAUDE.md to AGENTS.md, which is the thing Codex needs it to be called.
Step 2. Let the skill finish the job.
The import brings your files across. It does not convert everything, and the leftovers are the parts you built last: your shortcuts, subagents, hooks.
OpenAI published a skill for exactly that, and the funny part is where you install it.
It goes inside Claude Code:
npx -y skills add openai/skills --skill migrate-to-codex --agent claude-code
Save a copy of your Codex settings file before you run it. The skill rewrites that file from scratch, so anything you had set up in it goes.
Then run it and leave it alone. It converts what the import missed, checks its own work, and fixes what it got wrong. Ask it for a dual system and you keep both tools running side by side, which is what I did.
Step 3. Read your own rules back.
Neither of the first two steps does this for you.
They move your rules. Neither tells you what they say.
Mine had been growing for nine months and I could not have recited half of it. So open the folder and paste this:
I have just moved over from Claude Code. Go through this folder and make everything native to Codex.1. Read my instructions file, whatever it is called, and tell me in plain English what it asks you to do.2. Rewrite it so it reads as written for Codex, keeping every rule, and tell me which rules you changed the wording of and why.3. Find anything in here that only works in Claude Code, and name it.4. For each one, give me the Codex equivalent, or tell me plainly that there is not one.5. Change nothing else. List what you left alone.
It does not answer in the chat. It writes you a report.
Read it once and look for three things. The rules it reworded, and whether you agree with the new wording. Anything it says only works in Claude Code. And the list at the bottom of what it left alone.
Then I gave both tools the same task. Same brief, brand rules on one graphic. Claude Code made the one on the left and I was ready to bin it. Codex finished the one on the right.
It took seven versions and a lot of my feedback, so this is not a one-prompt result (nothing good ever is). It found a cause I could measure and fixed it.
The thing I did not expect was how it talks. Claude Code, explaining a fault to me:
Correct, and it fixed it. Now read that as someone who has never opened a terminal. Here is Codex, reviewing a graphic it had just made:
Same job. One is talking to an engineer. The other is talking to me.
And look at what it did first. It opened three of my past graphics and compared the new one against them, which took two minutes on a job I would have accepted at thirty seconds.
What everyone is (actually) arguing about
My own week is one data point, so I swept the last 30 days across Hacker News, Reddit, X, TikTok and Instagram. Every post about either tool, sorted by what it was actually about.
Almost every Claude Code thread is about the meter. Weekly caps, what people are spending, and not being able to see what it did with any of it.
Almost every Codex thread is somebody showing off what they built.
Get them to argue (with each other)
Claude Code can run as a command, not just a window you type into. Codex can run commands. So Codex can ask Claude a question and read the answer back, on its own, while it works.
I had Codex building a comparison graphic and I did not like it (I told it so, in those words). Instead of guessing, it went and asked Claude what was wrong with it.
Claude's diagnosis was right. The layout promised lots of differences, then filled itself with things both tools do. It also handed over three differences that were not real. Codex pushed back on all three, and Claude withdrew them:
Two labs arguing about my work, and the argument caught an invention that either of them would have shipped alone.
The thing I was (wrong) about
Letting an AI drive your screen has been the most disappointing thing in this whole category. Watching it hunt for a button you could have clicked in a second. I gave up on it a long time ago.
Then somebody uploaded a photo of himself, opened Canva, and gave both agents the same instruction: draw me.
So I pointed it at my own work. My infographic generator, the system behind every graphic you see on LinkedIn, and one line: “using only Computer Use, build a graphic from my recent newsletter.”
It opened Figma and did it. Read the newsletter, checked my design system (the one from last week), and designed the entire graphic!
That clip is sped up, but it honestly amazed me. It designed the whole thing from scratch and used Claude to self-critique the design until it hit 95 out of 100.
It did take a bit longer than I would have liked (a fair bit), so the Figma plugin is still what I run instead (not through the browser). But the accuracy is what’s impressive here, not the speed.
If you want to build the design system it was reading, last week’s newsletter is the place to go. And if you want the LinkedIn AI OS when it launches, fill this in.
There is one test (ARC-AGI 3) that measures this and it is not a coding test. It drops a model into a small abstract puzzle world it has never seen before, one turn at a time, with no instructions of any kind. It has to work out the rules by trying things, which is the same job as landing on a screen it has never met.
Astra got through levels using fewer moves than a human on 96% of them. On the same leaderboard, Claude Opus 5 sits at less than half its score.
Astra is the yellow line. Claude Opus 5 is the red dot below it. That lone yellow blob at the top is Astra again on a setup nobody else was given, which is the number you will see quoted everywhere.
The best setup is (both)
Do not pick one. The most effective way to use these is together, and OpenAI have made that the easy option. They built Codex into Claude Code as an add-on. Two lines, typed inside Claude Code:
/plugin marketplace add openai/codex-plugin-cc/plugin install codex@openai-codex
Now you can hand your work to Codex without leaving the tool you already know.
My graphics had been coming out worse, so I had the two of them go three rounds over my actual files and work out who should do what. This is what they agreed on:
Claude sets the structure, the hierarchy and the visual language.
Codex refines that same artwork, the spacing and diagrams.
They are both exceptional coding agents. Running one is a preference, running both is a division of labour.
I am not moving off Claude Code. I am running both, and letting the work decide.
That's it from me this week. See you again next week.
Stay curious, stay human and keep creating
— Charlie
