I Tracked What a React Native AI App Actually Costs to Build

The agency quotes are not lying — but they are hiding things. Here’s every line item, from twenty projects I watched up close.

Last winter, I watched a friend get three quotes for a React Native AI app. A voice-notes thing. Record meetings, get AI summaries, share to teammates. The quotes came back: $74,000. $112,000. $160,000. He chose the $74K one. Six months in, the agency was at 60% delivered and asking for another $40K. He shut it down.

Then he bought a $79 template.

That’s the story I keep watching repeat. And the reason it repeats is that nobody — not the agencies, not the calculators, not the “how much does an app cost in 2026” Medium posts — actually itemizes what’s inside the quote.

So I sat down and wrote out every single line item I’ve personally seen across maybe twenty indie React Native AI projects in the past year. I want to walk through it here, not as a sales pitch, but because the honest decomposition is shockingly hard to find online.

What’s inside the bracket

The agency line you’ve seen everywhere is “$60,000 to $150,000 for an app with AI/ML integration.” That bracket is mathematically correct and informationally useless. Inside it is a series of work blocks that almost nobody talks about by name.

There’s the design and information architecture phase, which solo devs always underestimate by a factor of four. It’s not making things pretty. It’s deciding what an “AI session” means in your domain — does it have a list view? a detail view? does it appear in search? what’s the empty state? what happens when the user is offline? That phase alone is sixty to a hundred hours.

There’s the auth and database layer. Supabase makes this fast, but “fast” is forty to eighty hours, not four. You’re writing schema migrations, designing row-level security policies that prove user A genuinely cannot read user B’s transcripts (which Apple’s reviewers will test), wiring email and OAuth, and putting your secret keys in edge functions instead of bundling them into the client.

There’s the AI integration itself, and this is where the bracket-quotes hide an enormous range. A calorie scanner — image goes in, structured nutrition data comes out — is maybe sixty to a hundred hours of provider work plus retry UX. A voice transcriber with streaming partial results is more like a hundred and sixty hours, because mobile audio is hard. PDF chat with retrieval-augmented generation is two hundred hours minimum, because you’re now also a vector-database engineer.

Then there’s the part that swallows half the budget on every project I’ve seen: the long tail of screens. Onboarding flows. History views. Settings. Paywalls. Empty states. Error states. Permission denials. The “what happens when the user has four thousand notes” scroll test. A real AI app is not one chat screen — it’s twenty-plus screens, each of which needs a dark mode, a tablet layout, and a VoiceOver label.

Add Stripe, push notifications, EAS Build configuration, App Store submission with its inevitable rejection, accessibility passes, and what looked like a $40K project quietly grew to $80K and seven months.

The recurring bill nobody mentions

Here’s the part that genuinely surprised me when I started running the numbers: the hours to build are not the most painful part. The painful part is that once you ship, you start paying.

GPT-4o Vision at roughly a cent per image. Whisper at six-tenths of a cent per minute of audio. Claude or GPT-4 for chat at three to fifteen dollars per million tokens. Supabase Pro at twenty-five a month. EAS at ninety-nine a month. Sentry. The Apple Developer Program. Maintenance — which the industry baselines at fifteen to twenty-five percent of initial development cost per year, every year.

A typical year-one all-in budget for a from-scratch React Native AI app I’ve watched ship: forty-five to seventy-eight thousand dollars. And that’s before you’ve spent a dollar on marketing.

Why templates are not the answer you think they are

I want to be careful here because “buy a template instead” is exactly the kind of pitch you’d expect from someone writing on a template company’s blog.

Most templates are scaffolds. They are not products. A free Expo boilerplate will save you the first forty hours and cost you the next two hundred, because what it actually gave you was a working router and a login screen — not an AI provider abstraction, not row-level security, not Stripe license grants, not a paywall, not the twenty other screens you need.

A paid UI-only kit gives you nicer login screens. You still own the entire backend, the AI wiring, and the licensing logic. That’s the 70% of hours nobody quotes for.

What’s worth paying for is architecture. Auth that already works. RLS policies someone has already designed. AI wiring abstract enough to swap providers.

Stripe license grants that already verify webhook signatures. Twenty-plus screens that have been designer-vetted and accessibility-checked. For about seventy-nine dollars on one of the Applighter AI templates — Voice Notes, Calorie Tracker, or Chat with PDF — that’s roughly what you’re buying. You bring your own AI API key, your own brand, your own twist on the idea. That’s two to six weeks of customization. Then you ship.

When building from scratch still makes sense

There are real cases. If you’re shipping a genuinely novel AI feature — on-device models with ExecuTorch, a custom-trained vision pipeline, a research-grade RAG architecture — no template covers it. Build it. If you have HIPAA or FedRAMP compliance constraints, a template’s RLS may not be enough. Build it. If you have an in-house React Native team that’s idle, building in-house may even be cheaper than waiting on a template’s release cycle.

But for the indie dev with an AI product hypothesis, the math is not subtle. Four months of your life testing an unvalidated idea, versus three weeks. Fifty thousand dollars of either your money or your time, versus seventy-nine dollars plus your AI bill.

I’m not saying don’t build. I’m saying: build the part that’s actually your product. Don’t rebuild auth, screens, and EAS configuration for the four hundredth time.

Read the full version — including the complete line-item table and the three real budget shapes (image-to-insight, voice-to-summary, document-to-chat) — on the Applighter blog.

👏 if you’ve watched this story play out too — and if you’ve shipped a React Native AI app, I’d love to hear your real number in the comments. The honest ones are hard to find.