OpenAI's Codex Can Build You a Full App in 23 Minutes. It Just Won't Let You Share It.

Codex Sites is OpenAI's new feature that lets ChatGPT's coding agent build, deploy, and host a working web app for you. OpenAI announced it on June 2, 2026 as part of a larger Codex update aimed squarely at non-developers. Not a mockup. A real app, with a database, on a live URL, built start to finish while I waited. It is OpenAI's answer to Lovable, Replit, and Google AI Studio: describe what you want, get a working thing.

I am a marketer, not an engineer, and I spent the morning putting it through real work on GPT-5.5 with medium reasoning. So treat the rest of this as a field test from a marketer's chair, and my honest feedback for other marketers: what it does well, the two walls I hit, and whether you should care today. Short version: a strong proof of concept, not something I'd build my work on yet. Someone who lives in code and doesn't blink at burning tokens might land somewhere completely different. I'm judging it on what marketing teams actually need.

I gave it one prompt. I asked for a site that generates four dad jokes a day, lets people vote on the best one, keeps the same four jokes for every visitor on a given day, saves the results, and holds them for a bracket-style championship later.

Then it worked for 23 minutes.

What came back was not a draft. It was a deployed, working app. It had a real backend database storing each day's jokes and every vote. It stopped people from voting twice. It saved the daily results so I could run that bracket later. It even ran its own checks, the kind a developer runs, before it called the job done.

I have used Claude Code enough to know the feeling of watching AI write working software. This was different in one specific way. With most coding tools, you get the code and then you deal with hosting, the part where non-developers get stuck. Codex Sites handled that part. The database, the deploy, the live URL, all of it, without me setting up a single account or server. For a marketer who can describe what they want but has never touched a deploy pipeline, that is the magic. You go from sentence to working app, and you skip the part that usually needs a developer.

So OpenAI built a website builder. Almost.

Once the site was live, I asked the obvious question. Can I share this? Can other people use it? Can I embed it on a web page?

My Codex said yes to all of it. It even handed me an iframe snippet to drop the site onto another page.

Then, a few minutes later, when I asked it to actually open the thing up so anyone could vote, it reversed itself. It told me it could not find a way to make the site truly public. Its words: "for truly public internet sharing, we'll need a Sites public access option or a different hosting path."

So I checked OpenAI's own documentation, because a tool being unsure about its own limits is worth verifying. The docs are clear. A Codex Site has exactly three sharing settings, and all three are walled inside your company. Admins only. Everyone in your workspace. Or specific people you name. There is no "anyone with the link" option. Every person who opens the site has to be a member of your OpenAI workspace, or sign in through an identity provider you set up first. There are no custom domains either. The site lives on an OpenAI-owned address you don't control.

My kids are not in my OpenAI workspace. They are kids.

That iframe snippet my Codex gave me? Embedding isn't documented anywhere I can find. I think it made it up. Which is its own small lesson: the AI was confidently wrong about the platform it runs on. Verify the limits before you build on them.

Here is the reframe. Codex Sites is not a website builder, the way Lovable or Replit are. It is an internal tool builder. It is built for the dashboard your ops team checks, the onboarding hub your new hires use, the idea board your department votes on. Work that lives inside the company. For that, the workspace-only wall isn't a bug. It's the point.

But the instinct, the second you build something fun, is to send it to someone. Right now you mostly can't. That gap, between what it feels like and what it is, is the thing to understand before you start.

For the record, my workaround: I open the site on my screen, and my kids come vote in person every morning. Which was the plan anyway. It fits my exact use case and almost nobody else's.

The second wall I hit mid-build. I asked it to add rate limiting and open the site up. It wrote the code, ran its checks, and then stopped cold with one line: "workspace is out of credits."

It had written the fix. It just couldn't ship it. I got one real round of changes in, and then I was done for a while.

Here's how the money works, because this is the part marketers need to understand before they get excited. OpenAI moved Codex to token-based pricing earlier this year. You get a small included allowance, and on a business plan that allowance is roughly the same as a single ChatGPT Plus seat. When you blow past it, Codex starts spending from a shared pool of credits your workspace has to fund. When that pool is empty, everything stops, including the deploy I was trying to run.

The allowance resets on a rolling five-hour clock. So the real cost isn't only dollars. It's rhythm. What could be a one-day build turns into this: work for a bit, hit the wall, wait five hours, work for a bit, wait five hours. I like to start a thing and finish it. This makes you ration a project across days. For a marketer trying to test an idea fast, that friction is the opposite of what you want.

There is a way to keep going. A business workspace can buy more credits and even set them to auto-refill, so Codex never hard-stops. I want to be accurate about that, because it's easy to assume you're simply locked out. You're not. But notice what that path is. It's a meter. You're paying per token, with no ceiling you set in advance, on a line item that's hard to forecast. There is no flat "give me more Codex for a fixed price" plan unless you go all the way to Enterprise and talk to a sales team.

OpenAI's other option is a Codex-only seat. It's usage-based, and here's the catch: it doesn't include ChatGPT at all. So you'd be choosing between the AI assistant your whole team uses every day and the building tool. I don't know many marketers who will give up ChatGPT to get a metered code seat.

Here's my honest opinion, labeled as opinion, not a prediction about anyone's business. Marketers, and the people who sign off on marketer budgets, love subscriptions with limits. We want to know the number before the month starts. Burning tokens with no predictable cap is how you get a finance conversation you don't want to have. If OpenAI wants marketing teams building seriously inside Codex, I think they'll eventually need a real subscription tier built for it. Predictable cost isn't a nice-to-have for us. It's how the purchase gets approved at all.

The thing that makes Codex Sites unusual is the same thing that limits it. Every comparable tool ships a public URL by default. Codex Sites keeps it inside your company.

One honest note on that table. Google AI Studio's quick "share" button looks public but isn't. It shares with other Google-account users and runs on your quota. You only get a truly public site by deploying it out to Cloud Run, an extra step. So don't lump its "share" in with Lovable's "publish." They are not the same thing.

I went deep on Sites because it's the piece a marketer feels first. But Sites shipped inside a bigger update, and the rest of it is pointed straight at people like us. So before you judge OpenAI Codex for marketers on one feature, here's the honest map of what else landed.

Start with the number that explains the whole thing. OpenAI says non-developers are now about one in five Codex users, and they're growing roughly three times faster than developers. Read that again. The fastest-growing group on a coding tool is people who don't code. Marketers, analysts, ops. That's why this update exists at all.

For marketers, the headline is role-specific plugins. OpenAI shipped six that bundle the apps, skills, and instructions for a job, so you're not starting from a blank prompt every time. The one built for us is the creative production plugin: hand it a brief and it helps turn that into reviewable assets, campaign boards, display-ad variations, and product or lifestyle shots, wired into tools we already use like Figma, Canva, and Shutterstock. A dedicated Marketing Strategy plugin is on the announced-but-not-shipped list, next to finance, consulting, and legal. That is a lot of new assets moving through a lot of tools, and none of it holds up if nobody outside Figma can find the finished version later. That is exactly the gap AI-ready digital asset management is built to close: one place those campaign boards and product shots land once they are approved.

I'll be straight with you: I haven't lived in the plugins the way I lived in Sites, so I won't hand you a verdict I didn't earn. What I can tell you is the direction is unmistakable. OpenAI has stopped treating marketers as visitors to a developer tool and started building lanes for us. And the same two walls from my Sites test, the workspace-only sharing and the metered cost, still apply to all of it, because those are platform choices, not Sites choices.

The deeper point is the one I keep coming back to. Marketers are becoming builders, and the tools are finally being built for that. Whether you start in Codex, Claude Code, or somewhere else, the skill that pays off is being able to describe what you want and judge what comes back. The repeatable parts of that work are worth saving as Claude skills, so the AI starts every job from your best thinking instead of a blank page.

If you want to build something real today, something public, something you can put on your own domain and send to a customer, I'd still reach for Claude Code and have someone set you up on GitHub and Vercel. That path gives you a public site and a predictable bill.

If you want an internal tool, a dashboard, a team voting board, a quick app only your coworkers need, and you're already on ChatGPT Business or Enterprise, Codex Sites is worth a real look right now. It is the fastest path from sentence to working app I've used, as long as "shareable" means "shareable with the people you work with."

For everything in between, it's a proof of concept. A very good one. The bones are here. I'd bet the walls come down over time, because they're choices, not limits of the technology. Public sharing, custom domains, a subscription that makes sense for a team. None of that is hard for OpenAI to add. It just isn't here yet.

To be fair to the tool, this is a marketer's verdict. I'm weighing predictable cost and public sharing because that's what my work needs. A developer who's fine burning tokens and building internal tools might love this on day one. I came at it as someone who wanted to ship something to the world this morning, and instead shipped something to my office.

Which my kids, to be clear, are voting on as I write this. The jokes are terrible. That part works perfectly.