---
title: "How to Build an App With ChatGPT (and Where It Stops)"
author: "Nate Laquis"
author_role: "Founder & CEO"
date: "2026-11-05"
category: "How to Build"
tags:
  - how to build an app with chatgpt
  - build an app with codex
  - ChatGPT app development
  - AI app builder
  - DIY app development
excerpt: "ChatGPT and Codex can get you a working prototype of a simple app for roughly $20 a month. This guide walks the steps, the real costs, and the point where a founder should stop and hand it to a team."
reading_time: "11 min read"
canonical_url: "https://kanopylabs.com/blog/how-to-build-an-app-with-chatgpt-and-where-it-stops"
---

# How to Build an App With ChatGPT (and Where It Stops)

## Short Answer: ChatGPT Can Build the Prototype, Not the Launch

You can build an app with ChatGPT by using Codex, OpenAI's coding agent inside your ChatGPT plan, to write the code from a written spec, run it on your own machine and fix what breaks. It works well for a single-purpose app with a handful of screens, and it gets you to a clickable prototype in days. It stops at the parts that happen outside the code: accounts with Apple and Google, payments, real user data, review, and everything that keeps the app alive after launch.

That split is the whole article. Below is the exact path, what it costs next to a hired build, and where I would tell a founder to stop doing it alone. If you are still deciding what to build at all, read our [MVP development guide](/blog/mvp-development-guide) first, because the best use of ChatGPT is on a product you have already cut down to one job.

![founder writing code on a laptop with an AI coding assistant](https://images.unsplash.com/photo-1517694712202-14dd9538aa97?w=800&q=80)

I checked OpenAI's own pricing and docs pages on November 5, 2026 for the facts below. Where a source did not say something, I say so rather than guess.

## What ChatGPT Can Build Today

Two different things get called "building an app with ChatGPT," and they behave very differently.

The first is chatting. You describe an app, ChatGPT writes code in the conversation, and you paste it into files yourself. That worked in 2023 and it still works, but it breaks down fast. Past a few files, the model loses track of what exists, and you become the copy-paste layer between it and your editor.

The second is Codex, and it is what you should use. OpenAI's Codex docs describe it as a way to start with a goal, idea or task, and list building a new feature, app or tool as one of four starting points. Codex runs in the ChatGPT desktop and mobile apps, on the web, in a command line tool, in an IDE extension, and as a cloud agent. It reads your project folder, edits many files at once, runs commands and reports back. That is the difference that matters. It works on the whole project instead of one pasted snippet.

A few other features show up on OpenAI's pages. ChatGPT now lists Work and an always-on agent called Dots on the Pro plan, plugins and apps inside ChatGPT, and a Sites feature for creating and publishing sites. I did not test Sites, and OpenAI's overview page does not describe what it hosts or what it costs to run at scale, so I would treat it as a way to publish a simple site, not as a home for an app with user accounts and payments. I also did not find canvas mentioned on the current pricing page, so I would not plan around it.

What you get in practice: generated screens, working navigation, forms, local data, calls to public APIs, and a lot of correct boilerplate. What you do not get is judgment about your product. The model builds what you describe. If the description is vague, you get a confident, polished version of the vague thing.

## What It Costs: Plans, Hosting and Store Fees

Here are the numbers I verified, so you can price the DIY path honestly.

- **ChatGPT Free:** $0 a month, with limited Codex for quick tasks.

- **ChatGPT Go:** $8 a month, still lightweight coding only.

- **ChatGPT Plus:** $20 a month. This is the first plan with Codex on the web, CLI, IDE extension and iOS, plus the cloud features. It is where most founders should start.

- **ChatGPT Pro:** from $100 a month, with tiers at $100, $200 and $500. You only need this if you hit usage limits every day.

- **Business:** $20 per user per month billed annually, or $25 billed monthly.

Limits matter more than the sticker price. OpenAI measures Plus usage in five-hour windows, and its own table gives ranges rather than promises: for example 15 to 160 local messages for one top-tier model and 350 to 3,000 for a lighter one. Cloud tasks use more of your allowance than local messages. When you run out, you buy credits. A founder building a real app over a month will feel these limits. Plan for Plus and a budget for extra credits, not for an all-you-can-eat plan.

Then the costs that are not ChatGPT:

- **Apple Developer Program:** $99 per membership year. Apple's own enrollment page says so, and notes fee waivers for qualifying nonprofits, schools and government bodies.

- **Google Play:** a one-time $25 registration fee, per Google's help page.

- **Hosting:** Vercel's free Hobby plan is $0 but is for personal, non-commercial use only, per Vercel's pricing page. The Pro plan is $20 a month. A real product with paying users needs the paid tier or an equivalent.

- **Mobile builds:** Expo's EAS service has a free tier with 15 Android and 15 iOS builds a month on a slow queue, a $19 a month Starter plan, and a $199 a month Production plan.

Add it up for a founder who builds a simple mobile app on Plus and ships to both stores: about $20 a month for ChatGPT, $124 in store fees in year one, and $0 to $40 a month for hosting and builds. Call it $400 to $800 in the first year, plus whatever the backend services you add cost. That is the cash number. The real cost is your time, and it is the part founders underestimate.

A hired build is a different order of magnitude. Market ranges for a first version of a mobile app sit roughly between $30,000 and $150,000 depending on features, platforms and who builds it. Our full breakdown, with what drives each number, is on the [app development cost](/app-development-cost) page.

## How to Build an App With ChatGPT, Step by Step

This is the order I would follow. It is written for a founder who has never shipped code.

### Step 1: Write a one-page spec before you open ChatGPT

Write who the app is for, the one thing they do in it, and the four to six screens that make that thing possible. List what is out. Codex is only as good as this page. A spec that says "a social app for runners" gets you a generic feed. A spec that says "a runner logs a run in under ten seconds and sees a weekly total" gets you something you can test.

### Step 2: Pick the simplest stack the model knows well

Ask for React Native with Expo if you want iOS and Android from one codebase, or a plain Next.js web app if a website will do. Both are heavily represented in public code, so the model makes fewer mistakes. Avoid exotic frameworks. A common stack is a boring stack, and boring is what you want when you cannot read the code yourself.

### Step 3: Set up the project and install Codex

Install Node, create an empty project folder, and open it with Codex through the desktop app, the command line tool or the IDE extension. Put your one-page spec in the folder as a plain text file so Codex can reread it. Ask it to scaffold the project and run it.

### Step 4: Build one screen at a time

Do not ask for the whole app. Ask for the first screen, run it, look at it, then ask for the next. After each one, commit to Git so you can roll back. This is the habit that separates a founder who finishes from one who ends up with an app nobody, including the model, can untangle.

### Step 5: Use a real backend for anything that must persist

Local state is fine for a demo. The moment two people or two devices need the same data, you need a database and sign-in. Hosted services such as Supabase or Firebase are the usual picks, and Codex can wire them in. Keep the keys out of the code, and have the model put them in environment files.

### Step 6: Test on a real phone

Run the app on your own device with Expo Go, then share a TestFlight build with five people. Watch them use it. The bugs you find will be about confusing flows more than broken code.

### Step 7: Ask Codex to review its own work

Before you show anyone, ask it to list security risks, missing error states and anything that would fail Apple review. It will find real issues. It will not find all of them, which is the subject of the next section.

![person testing a mobile app prototype on a smartphone](https://images.unsplash.com/photo-1512941937669-90a1b58e7e9c?w=800&q=80)

## Where It Stops: The Failure Points

Here is where I see founders get stuck, in the order they hit it.

### Context and drift

OpenAI's pricing page lists context windows in the tens to hundreds of thousands of tokens depending on the plan and mode, for example 54K for instant replies and 256K for reasoning on Plus. That sounds large. A mid-sized app plus its history of changes outgrows it. Past that point the model forgets earlier decisions, renames things, duplicates logic and fixes one bug by reintroducing another. You notice it as "every fix breaks something else."

### Authentication and user data

Sign-in looks easy, and the code it writes usually runs. The failures are quiet. Tokens stored where they should not be, screens that show another user's data when you change an ID, password reset flows that leak whether an email exists. Apple also requires that an app which lets people create an account must let them delete it from inside the app, and requires a privacy policy link in the app and in App Store Connect. Both are in Apple's App Review Guidelines, and neither is something the model volunteers unless you ask. Our [guide to taking vibe-coded apps to production quality](/blog/vibe-coding-to-production-quality-guide) covers the vulnerability patterns in depth, so I will not repeat it here.

### Payments

If you sell digital features or subscriptions inside an iOS app, Apple's guidelines require in-app purchase, not your own payment link. That means StoreKit or a service such as RevenueCat, product setup in App Store Connect, sandbox testing and receipt handling. Codex can write the client code. It cannot set up your tax and banking details with Apple, and it cannot test a purchase flow against your real account for you. This is where many DIY apps lose their first weekend, then their second.

### Store accounts and review

The $99 and $25 fees are the easy part. Apple's guideline 4.2 says an app has to offer more than a repackaged website, and rejections for thin functionality are common. Google's rules for new personal developer accounts require a closed test with 12 testers who stay opted in for 14 days before you can apply for production access. So your first Android release has a built-in two-week wait, and you need twelve real people.

### Deployment and everything after

Codex can run in a cloud sandbox and open code changes for you. It does not become your operations team. Someone has to own the production database, backups, the error alerts at 2 a.m., the OS update that breaks your camera screen in September, and the user who emails that they were charged twice. If that someone is you, budget the hours.

![laptop screen showing app analytics and error monitoring dashboards](https://images.unsplash.com/photo-1460925895917-afdab827c52f?w=800&q=80)

## DIY With ChatGPT vs a Hired Build: Real Ranges

Here is the comparison I would show a founder across a table. The hired numbers are market ranges, not a quote from us.

- **Cash cost, DIY:** roughly $400 to $800 in year one for a simple app, as worked out above. Add backend services and your own labour.

- **Cash cost, hired:** roughly $30,000 to $150,000 for a first version, driven mostly by number of screens, integrations and whether you need design from scratch. See the [app development cost](/app-development-cost) breakdown.

- **Time to a prototype, DIY:** a few days to two weeks for a founder who already has a clear spec.

- **Time to the App Store, DIY:** most first-timers I talk to land at two to four months, because review rejections, payment setup and the Google test period all add calendar time that no tool shortens.

- **Time to the App Store, hired:** our [timeline guide](/blog/app-development-timeline-how-long-does-it-take) puts an MVP or simple app at 6 to 12 weeks, and mid-complexity apps at 3 to 6 months.

The DIY number is small because your time is free on the spreadsheet. It is not free. If you are a founder whose time is better spent on customers, fundraising or the day job that pays rent, three months of evenings is the most expensive line in the budget.

The reverse is also true. A hired build before you know anyone wants the product is the wrong spend. A $60,000 build of an untested idea is a worse outcome than a $600 prototype that tells you in two weeks that nobody cares. So the right comparison is not DIY against hired. It is a cheap test against a real build, in that order.

## What to Build First With ChatGPT

Use the cheap tool on the cheap questions. These are the things I would build with ChatGPT before spending real money.

1. **A clickable prototype of the core flow.** Three to five screens, fake data, no accounts. Put it in front of ten target users. You are testing whether the idea makes sense, not whether the code is good.

2. **A single-purpose web tool.** One input, one result. No sign-in, no payments. If strangers use it, you learned something.

3. **A landing page with a waitlist.** Codex writes this well, and it is the cheapest demand test there is.

4. **Internal tools.** A dashboard only your team uses. The risk of a bug is low because the users are you.

5. **A throwaway version of the riskiest feature.** If the whole product hangs on one thing working, such as photo recognition or a pricing calculation, build only that and see if it holds.

And what I would not build with it alone: anything that stores other people's sensitive data, moves their money, or has to pass App Store review under your name. You can start those with ChatGPT. You should not finish them without someone who has shipped them before.

A useful example of scope done right is Fake Friends, an Instagram unfollower tool we designed and built. It does one job, tells you who unfollowed you from Instagram's own data export, and it went from first commit to production in two days with live payments, sign-in and a custom domain. The speed came from a narrow scope and a clear rule about what the product would never do, which was ask for a password. You can read the full build on the [Fake Friends case study](/work/fakefriends). The lesson for a ChatGPT build is the same: a smaller product ships, and a bigger one stalls.

## When to Stop and Bring in a Team

Switch from DIY to a hired build when you can answer yes to any of these.

- People you do not know have used the prototype and asked when they can pay for it.

- The next feature on your list involves money, accounts or private data.

- You have spent more than a month and the app keeps breaking in new places after each fix.

- You have a launch date, an investor conversation or a partner waiting.

- You need iOS and Android both, working well, in the stores, with real support behind them.

The good news is that the ChatGPT work is not wasted. A tested prototype and a written spec are what make a hired build cheaper and faster, because the team spends its first weeks building instead of guessing. Bring the prototype, the notes from your ten user tests and the list of things you cut. Most of the scoping work is already done.

If you are about to build a mobile app, the stack, the screens and the store setup are the part where a short conversation saves months. Kanopy builds iOS and Android apps for founders, and we are glad to look at what you have made with ChatGPT and tell you honestly what to keep, what to rebuild and what to cut. Tell us what you are planning and [Book a free strategy call](/get-started).

---

*Originally published on [Kanopy Labs](https://kanopylabs.com/blog/how-to-build-an-app-with-chatgpt-and-where-it-stops)*