Codex or Claude Code? Same Task, Head to Head, and It's Not Close

Way too many people have been asking me which to pick, Codex or Claude Code, so today I’m answering it head-on.

My answer is simple: if you can reliably run official-subscription Claude Code, there is no reason at all to use Codex. Or, put another way, you can wire Codex in as a code-audit helper for Claude Code, and that’s it. Beyond that there’s really no reason to go out of your way to use Codex.

What makes the comparison fair: same config, same skill, same prompt

Stating a conclusion is boring, so I took a workflow I run all the time and used it to put the gap between these two in front of you.

The key is that the variables have to match: same config, same skill, same prompt. The task was to research “how to automatically write a WeChat Official Account article from a piece of content, auto-generate the images, and push it into the Official Account’s drafts box.” Then present the research report as a web page in the PPVI light theme and save it to a folder on the D drive. (On the Codex side I used the fast variant of GPT-5.5 with thinking depth cranked all the way up.)

Codex’s result: at first glance, not bad

Start with what Codex produced. At first glance it’s fine: there’s a one-sentence verdict, four pipeline steps, some caveats, and a recommended path.

If you hadn’t seen what Claude Code produced, you’d probably think: this is pretty good, better than researching it myself, and it gets the facts right—even if it isn’t very detailed.

Claude Code’s result: same prompt, and you can tell what’s solid and what’s junk

But once you see the page Claude Code produced under the same prompt and the same skill, you can judge for yourself what’s solid and what’s junk.

The research report it produced has:

• first, a clickable table of contents;
• a conclusion-first summary, technical feasibility, framework principles, default strategy;
• the whole panoramic architecture data flow—the design polish alone is in a different league;
• pros and cons of the two technical routes;
• the hard account-eligibility gate: what the difference is between a subscription account and a service account, and what the APIs look like;
• exactly how the four API gates on the drafts path work;
• the length and format limits on each field;
• what framework to use for automatic content writing and how to build the article structure;
• the truth about Official Account layout—an Official Account post isn’t a Markdown txt file; it carries tags and styles, and the report shows how to produce that layout;
• specs, dimensions, and the illustration pipeline for auto-generated images and covers;
• a head-to-head of image models: Nano Banana Pro or Seedream, their respective prices, their strengths and weaknesses for Official Account imagery, what goes wrong with covers that contain text, and all the traps in overlaying Chinese type with Pillow—it anticipated all of that and wrote it in;
• image post-processing, off-the-shelf tools and the open-source landscape (every open-source SDK listed), implementation advice, risks and red lines;
• and finally an honest boundary section: what has been verified, what still needs checking, and the sources.

Codex has these sections too, but put the two side by side and it’s not close.

I’ve put this Claude Code report in the downloads section on this page. It’s both the answer sheet for this comparison and a hands-on manual for Official Account automation you can follow directly.

Why do so many people think Codex is better? Usually they got burned by a “fake Claude Code”

I get where a lot of creators are coming from: they can’t get an official subscription to Claude Code, so they run a third-party relay API instead. A “relay Claude Code” runs into all sorts of problems—tool use breaking, context getting polluted.

And yes, relay Claude Code may well be less usable than Codex. But here’s the issue: what made them start saying Codex is great at this and great at that was a “fake” Claude Code. I find that genuinely misleading. Take a version a relay has hollowed out, call it Claude Code, and the conclusion is skewed before you start.

So the premise of this comparison is the real official-subscription Claude Code. Align the variables and the gap becomes visible.

Tomorrow’s preview: the rumors about the July 8 “account bans”

There’s another thing going around hard right now, with plenty of creators saying that on July 8 Anthropic will ban accounts en masse and start “scanning faces.”

I’ll do a separate episode on that tomorrow: how everyone can get their data backed up before July 8, and what you can do to keep yourself out of that machinery as much as possible.

That’s the share for today. See you tomorrow.

Source: EP0047_audio.mp3 · ASR model gemini-2.5-pro (chunked parallel) · full text of the original recording

[00:00] Alright, we’re finally doing this topic: Codex or Claude Code, which one do you pick? My answer here is dead simple. If you can reliably use Claude Code on the official subscription, then there’s really no need to use Codex—or you can wire Codex in as a code-audit assistant for Claude Code, and that’s it. There’s absolutely no need to use Codex. So many people have been asking me lately that I’m going to take something I do all the time and run the two of them against each other, and show you just how big the gap between these two really is.

[00:26] First, I used the same config, the same skill, the same prompt. I had it research how to automatically write WeChat Official Account articles based on content, generate the images, and push them into the account’s drafts. For the research report I told it to lay it out in PPVI’s light web-page style and save it to a folder on my D drive. It ran on the high-speed version of 5.5 with ultra-high thinking depth. Take a look at what it produced. This is Codex’s result.

[00:51] At first glance it’s not bad, right? There’s a one-line verdict, then four pipeline steps, then some things to watch out for, a recommended path. Let’s keep scrolling. This is already a front-end task where it was handed a skill. And this is the kind of thing it writes. Then right after, it gives some practical tips, then some approaches to automatic illustration, including the points most likely to fail. It gives the implementation order too.

[01:16] So it’s roughly this kind of page. But before you’ve seen what Claude Code produced, you might think, hey, this is pretty good, right? More accurate than researching it myself, right. It’s not very detailed, but—then you see the page Claude Code produced. Same prompt, same skill, this page. What’s solid and what’s flimsy, judge for yourself. “Task complete, please advise.” This is the research report Claude Code produced.

[01:41] First off, this one has a table of contents you can click to jump around. Then a conclusion-first summary, the technical feasibility, then its framework principles and default strategy, the whole big-picture flow of the framework. Look at the sense of design in this front end—no comparison. Then two technical routes, the pros and cons of the two routes. The hard requirements on account credentials. What the difference is between a subscription account and a service account.

[02:07] What its APIs look like. The whole draft-box path needs four API gates—what exactly is going on there, right. Which fields, what character limits they have, or what format limits. What kind of framework you can use for automatic writing. What the article structure looks like. WeChat Official Account formatting, right—it isn’t just a Markdown, a txt kind of thing, it actually has tags, it has styles. So how do you pull that formatting off. Automatic illustration and cover images, cover specs,

[02:32] dimensions, an image pipeline—it’s all in here. Then a head-to-head on image models: do you use Nano Banana Pro or Seedream 4.0, right, with their matching prices. For generating images for a WeChat Official Account, what each one’s strongest suit and weakest point is. What problems you’ll hit with covers that have text on them, and what solutions there are. Stacking Chinese text with Pillow has its own traps—it already anticipated all of that, saw it and put it right here. Then image post-processing, off-the-shelf tools and the whole open-source landscape,

[02:57] open-source SDKs, all listed out here for you. Then implementation advice, risks and red lines, honest boundaries—which parts it’s already verified and which bits are still to be confirmed. Sources. Codex has those too, granted. To me the winner is obvious. So a lot of creators—and I can understand it—have no way to get onto Claude Code’s official subscription. They use a third-party relay’s API and run into all sorts of tool-use problems, like context pollution.

[03:23] It’s true that Claude Code over a relay may not even be as usable as Codex. But then, just because they used what’s basically a fake Claude Code, they start saying Codex is great at this and great at that—I think that’s leading people astray. And there’s been a lot of wild talk lately, a lot of creators saying that on July 8 Anthropic and Claude are going to ban accounts en masse, right, that they’re going to scan your face. I’ll do a whole separate episode on that topic tomorrow: how everyone should get their data backed up before July 8, and what options there are

[03:49] that might keep you out of that face-scan mechanism. That’s it for today. See you tomorrow, bye-bye.