AI mobile app development in 2026: what AI does at each stage, the tools, and the costs

It is mobile app development where an AI coding agent does the writing. You describe a feature in plain language, the agent plans it, writes the code in the platform’s language (Swift for iPhone, Kotlin for Android, or JavaScript for React Native), builds the app and runs it in a simulator or on a device. You review the plan and the result and decide what ships.

Three things did not change. The App Store and Google Play accounts are in your name, and so is the responsibility for what the app does. Apple’s and Google’s review rules apply to an AI-written app exactly as to any other. And the agent writes what the description says: details the description leaves out, the agent guesses.

The table below goes through the seven stages of a mobile app. In short: AI now does most of the writing, building and first-pass testing. You still own the decisions, the accounts, the real-device check and the release.

Spec. The agent can turn a rough description into a structured plan with screens, states and edge cases. The decisions in it (who the user is, what the first version leaves out) are yours. The one-page requirements template is a starting point.

Design. Agents can build screens from a description, a screenshot or a design file, using the platform’s own components. Taste and the decision of what is on each screen stay with you.

Code. Xcode 26.3 (February 2026) supports “Anthropic’s Claude Agent and OpenAI’s Codex”, and Apple says agents “can search documentation, explore file structures, update project settings, and verify their work visually by capturing Xcode Previews.” In Android Studio, Gemini’s Agent Mode plans a task you describe and makes “changes across multiple files”.

Test. Agents can now run the app and look at it. Google’s documentation: “The AI agent in Android Studio has access to tools to deploy an app to the connected device, inspect what is shown on the screen, take screenshots, check Logcat for errors, and interact with the running app through adb shell input.” How AI agents test iOS apps in the simulator covers the iPhone side.

Backend. Agents connect the app to a backend such as Supabase or Firebase and write the code that calls it. Which backend you use, and who holds the keys, is your decision. Keys belong in configuration, not in the chat: how to keep API keys out of an AI agent’s chat.

Release. The agent can prepare the build. Uploading it, the store listing, the privacy answers and the reply to App Review are done in your accounts, by you. How to publish an app on the App Store walks the Apple side.

Maintenance. Every new feature is another round of spec, code and test. Apps built without a structure tend to break older features when new ones are added; why a vibe-coded app breaks with every new feature explains why and what prevents it.

Four kinds, sorted by where the agent runs and what code you get.

1. Agents inside the platform’s own tools. Xcode for iPhone apps and Android Studio for Android apps both have an agent built in now. Xcode 26.3 “makes its capabilities available through the Model Context Protocol”, so compatible agents can work with it; Android Studio’s Agent Mode runs on Gemini and asks you to “review and approve any changes as the agent works.” These suit developers who already work in one of the two tools, one platform at a time.

2. General coding agents. Claude Code, Codex and Cursor are pointed at a project folder and write any code, including Swift and Kotlin. They are not specific to mobile; you set up the project, the simulator and the build yourself. Claude Code with Xcode shows that setup for iPhone apps.

3. Hosted AI app builders. You describe the app in a browser and the builder writes and builds it on its own servers, usually paid in credits. What they produce differs a lot. Several write a React Native (Expo) project that runs on both phones. Rork writes native apps: “Rork writes a real Android Studio project in Kotlin”, and “each app has its own code”; code export is “available to paid users only.” Some put a web app into the stores: Base44’s docs say “Your mobile app runs your published Base44 app inside a secure web view.”

4. Desktop builders that run your own agent. The builder runs on your computer and uses your own AI subscription, so there are no credits. Modaal is one: a Mac app that builds a native SwiftUI app for iPhone and a native Jetpack Compose app for Android from one project, on your Claude or Cursor subscription.

How to choose an AI mobile app builder compares the builders by stack, and build a mobile app with AI: every route maps the six routes by where you start.

The tool decides the code, so decide the code first. There are three options:

Native: Swift with SwiftUI for iPhone, Kotlin with Jetpack Compose for Android. The app uses each platform’s own components and gets new platform features, such as new layouts for new devices, as soon as Apple or Google ships them.

React Native or Flutter: one codebase draws the screens on both phones. It is a good fit when both apps should look the same and your team writes JavaScript or Dart.

A web app in a native shell: your website, packaged for both stores. It is the fastest way to get a web app listed, with the limits of a web view.

Changing the stack later means rebuilding the app. Bolt’s own guide says it plainly: “Projects created for web do not easily switch over to mobile.” React Native vs native goes through the trade-offs.

Yes, in four ways: one codebase for both phones (React Native or Flutter), one shared core with two native interfaces, two separate native apps, or a web app wrapped for both stores. Each gives you a different pair of apps. Can AI build iOS and Android apps at the same time? lists ten builders by which way they use.

Features that depend on each other. An agent writes one feature well. Problems start when the fifth feature changes something the second one relied on. A written spec per feature, and tests that run on every change, are what keep older features working.

Platform states the agent never sees. Rotation, dark mode, large text, a lost connection, a call that interrupts the app, and on foldables every pose. An agent tests the states it is asked to test, so list them in the spec.

Secrets in the conversation. API keys pasted into the chat end up in the agent’s context and, often, in the code. Keep them in configuration files the agent references but does not print.

The store steps. No tool removes these. Apple charges “99 USD per membership year” for the Developer Program; Google Play charges a one-time 25 USD registration. New personal Google Play accounts must also “run a closed test for their app with a minimum of 12 testers who have been opted in continuously for at least 14 days” before the app can go to production. Plan that into the release date.

Three costs, and only the third depends on the tool.

The store accounts: 99 USD per year for Apple, 25 USD once for Google Play.

The backend: most backends have a free tier for a first version; the bill grows with users, not with the build.

The AI: tools charge in one of two ways. Hosted builders mostly sell credits, so each prompt and each fix costs credits. Tools that run your own agent use the Claude, ChatGPT or Cursor subscription you already pay for, so the cost per month is fixed. Modaal is free for one project with unlimited prompts on one platform; both platforms are on the Pro plan, €9 per user per month billed annually or €15 monthly, see pricing.

For the full calculation, including what a freelancer or an agency adds, see how much does it cost to create a mobile app.

That is the second meaning of the term, and the first decision is where the model runs.

On the device: both Apple and Google now let apps use models that run on the phone. There is no cost per request, it works offline, and the user’s data stays on the phone.

In the cloud: an API call to a model provider. You get the most capable models, but every request is billed, and the app needs a network connection.

An AI agent can build either kind of feature into a native app. The product questions (what the feature does for the user, what it costs per active user, what happens when the model is wrong) are covered in how to add AI features to a mobile app.

Day 1: write one page. Who the user is, the one job the first version does, and every way that job can end.

Day 2: pick the stack. Native, React Native or a web shell, using the three options above. Then pick a tool that writes that stack.

Day 3: open the store accounts. Apple and Google, in your name or your company’s. Google’s closed test needs 12 testers for 14 days before production, so the account comes first.

Days 4–6: build one feature, on a real phone. One feature, all its endings, tested in the simulator by the agent and on your phone by you.

Day 7: read the code you own. Open the project in Xcode or Android Studio and confirm it builds without the tool. If it does not, the app depends on that tool for every future build.