App Development
Last reviewed
May 13, 2026
Sources
80 citations
Review status
Source-backed
Revision
v2 ยท 5,777 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 13, 2026
Sources
80 citations
Review status
Source-backed
Revision
v2 ยท 5,777 words
Add missing citations, update stale details, or suggest a clearer explanation.
See also: App Development ChatGPT Plugins
AI in app development refers to the use of large language models, generative AI, and AI agents to design, build, test, and ship software applications, with a particular focus on consumer apps, mobile apps, and web products rather than general-purpose programming. The field grew quickly after the launch of GPT-4 in March 2023 and accelerated when Anthropic released Claude 3.5 Sonnet in June 2024, which was the first model capable of writing complete, working applications from a single natural language prompt with high reliability.
This article covers AI-native "text-to-app" builders such as Bolt.new, Lovable, Replit Agent, and Vercel v0; AI features inside no-code and low-code platforms; AI features inside mobile IDEs such as Xcode and Android Studio; AI-augmented app store review and discovery on Google Play and the Apple App Store; AI backend platforms and SDKs; and the marketplaces where AI-generated apps are now sold and shared. The term "vibe coding", coined by Andrej Karpathy in February 2025, is now widely used to describe the practice of building apps by describing them in natural language and accepting the AI's output with minimal review.
App development is distinct from general software development because it centers on shipping a finished product, often to an app store, with packaging, signing, distribution, metadata, and post-launch metrics treated as first-class concerns. AI is now embedded at every stage of that pipeline.
The AI app development stack in 2025 looks very different from the no-code stack of 2020. A typical AI-built consumer app now passes through several AI layers before it reaches a user:
Each step removes a traditional barrier (front-end coding, server setup, store copywriting), and each step introduces a new failure mode (exposed API keys, missing row-level security, inference cost blowups, AI slop in store listings).
Before generative AI, the dominant tools for non-developers building apps were visual builders such as Bubble (founded 2012), Webflow (2013), Glide (2018), and FlutterFlow (2020). These platforms used drag-and-drop canvases and predefined components. Building an app required no code, but it did require a user to understand layouts, data models, and conditional logic. Microsoft launched Power Apps in 2016 as the enterprise version of the same idea.
GitHub Copilot reached general availability in June 2022 and brought autocomplete-style code generation into editors. The launch of ChatGPT in November 2022 was the larger catalyst because it allowed non-developers to describe what they wanted in plain English and get working code back. In June 2023, Anton Osika at Depict.ai wrote GPT Engineer over a weekend, a command-line tool that turned a natural language brief into a small codebase. GPT Engineer became the most-starred new project on GitHub in days.
In September 2023, Vercel announced v0, an AI tool that generated React and Tailwind UI components from text prompts. The public beta opened in October 2023 and signed up 100,000 people in the first three weeks. Microsoft announced Power Apps Copilot in March 2023, letting users build line-of-business apps by describing them in plain language.
2024 was the breakout year. The release of Claude 3.5 Sonnet in June 2024 was the inflection point because it was the first model that could reliably produce a complete, working application in a single zero-shot pass. Three launches followed within months:
Claude Artifacts reached general availability in August 2024 and let any Claude user create small interactive web apps in a side panel.
On February 2, 2025, Andrej Karpathy posted on X that he had been doing "vibe coding", a style where you describe what you want, accept what the model gives you, and never read the code. The term took off. Merriam-Webster added it to its slang tracker in March, and Collins English Dictionary named it Word of the Year for 2025.
Growth in the category in 2025 was unusual even by AI standards. Bolt.new reached $20 million in annual recurring revenue (ARR) within two months of launch. Lovable hit $10 million ARR in 60 days, $100 million ARR by July 2025, and $200 million ARR by November 2025. Replit's ARR grew from $2.8 million to roughly $150 million in less than a year, driven mostly by Agent usage.
The app stores felt the result. Apple's App Store received 557,000 new app submissions in 2025, a 24% jump on the previous year and the largest single-year wave since 2016. Review times for new submissions stretched to 14 to 45 days.
Text-to-app builders accept a natural language description and produce a complete, runnable app, usually a full-stack web app with a frontend, a database, and authentication. Most use a single foundation model (Claude or GPT-4 class) plus a custom orchestrator that handles file edits, dependency installation, and live previews.
| Tool | Launched | Parent / company | Notable feature |
|---|---|---|---|
| Vercel v0 | October 2023 (beta) | Vercel | First mainstream generative UI tool, React + Tailwind + shadcn/ui output |
| Replit Agent | September 2024 | Replit | Agentic full-stack builder inside the Replit cloud IDE |
| Bolt.new | October 3, 2024 | StackBlitz | Runs the whole dev environment in the browser via WebContainers |
| Lovable | November 2024 | Lovable AB (Stockholm) | Conversational app builder, Supabase by default |
| Create.xyz (Anything) | 2024 | Create | Natural language prompts plus a visual code-aware editor |
| Codev | 2024 | Codev, Y Combinator | Next.js plus Supabase template, full code export |
| Base44 | January 2025 | Acquired by Wix, June 2025 | Six-month-old solo project sold for $80 million in cash |
| Replit Mobile Apps | February 2025 | Replit | React Native and Expo support, three-click App Store publish |
| v0.app | August 2025 (rebrand) | Vercel | Full-stack agent, in-product PRs and previews |
| Cursor Composer | October 2025 | Cursor (Anysphere) | Mixture-of-experts model trained with reinforcement learning on real codebases |
v0 started in September 2023 as a generative UI tool. The first version generated React components styled with Tailwind CSS and the shadcn/ui library, with each component shipped as paste-ready code. The public beta opened in October 2023 and hit 100,000 waitlist signups in three weeks. By February 2025, v0 was estimated to generate roughly $42 million in ARR, around 21% of Vercel's total revenue.
v0 evolved into a full-stack agent in 2025. The August 2025 relaunch (v0.app) added a sandbox runtime, GitHub repo imports, environment variables pulled from Vercel projects, and a Git panel that creates a branch per chat and opens pull requests against main. By February 2026, Vercel reported more than 4 million users had tried v0, up from 3.5 million at the time of its Series F in September 2025.
Replit had been a browser IDE since 2016. Replit Agent, released in September 2024, turned the IDE into a builder that planned and executed tasks. Agent v2 launched in early 2025 powered by Claude 3.7 Sonnet, with improvements in code quality and speed. In February 2025, Replit added React Native and Expo support, letting users build native iOS and Android apps by chatting, then publish them to the App Store and Google Play with three clicks. The official Replit mobile app reached number 1 on the App Store in Developer Tools that year.
Replit switched from a flat per-checkpoint price to effort-based pricing in mid-2025. Simple changes cost as little as $0.06; complex builds can cost several dollars per checkpoint. The change caused public complaints from a part of the user base, who said the model was harder to predict than the old $0.25 per checkpoint rule. Replit closed a $250 million round in September 2025 at a $3 billion valuation, led by Prysm Capital with participation from Amex Ventures and Google's AI Futures Fund.
StackBlitz had been working on WebContainers, a technology that runs Node.js inside the browser, for years. The team paired WebContainers with Claude 3.5 Sonnet and launched Bolt.new on October 3, 2024 with one tweet from CEO Eric Simons. The numbers that followed are now widely cited as the fastest software ramp on record:
Bolt's pitch was that it ran the full development stack in the browser. There was no install, no terminal, and no local dependency setup. Users typed a prompt, Bolt generated files and ran them inside a WebContainer, and the user could deploy to Netlify with one click. The platform later added Supabase integrations for authentication and storage.
Lovable, founded by Anton Osika and Fabian Hedin in Stockholm, is the commercial successor to GPT Engineer. Two earlier products under the GPT Engineer App brand launched in 2024 and failed. The team rebranded to Lovable in November 2024 and relaunched. The new product hit number 1 on Product Hunt and Hacker News on the same day.
Growth from there was the fastest in the history of software. By month two, Lovable was at $10 million ARR. By month seven (July 2025), Accel led a $200 million Series A at a $1.8 billion valuation, making Lovable a unicorn. By November 2025, ARR was $200 million. In December 2025, Lovable raised again at a $6.6 billion valuation. The company has stayed in Stockholm, has fewer than 100 employees as of November 2025, and uses no large sales organization.
Lovable's product is a chat interface that builds React + Supabase apps. It defaults to publishing on a lovable.app subdomain, with one-click deploy to a custom domain and GitHub sync.
Base44 was founded by Maor Shlomo in Israel in January 2025 as a side project. It used the same playbook as Lovable, with a chat-based builder that handled databases, authentication, and deployment under the hood. Wix acquired Base44 in June 2025 for $80 million in cash plus earn-outs, with the team still at six people and no outside funding. Wix moved Base44's technology into its main platform as an AI app builder.
Cursor, an AI-first code editor from Anysphere, sits one step closer to traditional development than the pure text-to-app builders. Cursor 2.0 (October 2025) introduced Composer, a mixture-of-experts model trained with reinforcement learning on real codebases. Composer edits multiple files at once, runs tests, fixes lint errors, and navigates large projects. Cursor 2.0 also added an embedded browser tool for agent-driven UI testing, voice input, and a parallel agent system that can run up to eight agents at once.
The established no-code platforms added AI features in 2023 and 2024, partly to keep up with the new wave of text-to-app builders.
Bubble.io added an AI App Generator that takes a text prompt and produces a working app, often a database-backed web app, in about five to seven minutes. The user can then switch between the AI agent and Bubble's visual editor for fine-grained control. Bubble launched its native mobile app builder in August 2025, letting users publish iOS and Android apps directly.
Webflow introduced its AI Assistant in October 2024. It generates styled sections (pricing, features, testimonials), produces CMS content, and answers in-product help questions without leaving the canvas. In February 2025, Webflow added an AI Site Builder beta that scaffolds an entire site from a prompt. AI code components let users describe interactive features like pricing calculators and have Webflow build them.
FlutterFlow is a visual builder for Flutter cross-platform apps. It integrated Gemini as a first-party action so makers can call Gemini for text generation, image analysis, and tool calls without writing Dart. FlutterFlow's AI Agent feature lets a builder pick a provider (Google, OpenAI, or Anthropic) and embed a chatbot or assistant directly into the app's UI.
Glide started as a way to turn Google Sheets into mobile apps. Glide AI lets a user type a description like "inventory tracker for my warehouse" and have the platform generate the database schema, screens, user roles, and sample data. According to Glide, the creator typically produces a working prototype in under 15 minutes.
Airtable rolled out Field Agents (AI-powered fields that process data as it lands) and Omni (a conversational sidekick that builds apps, interfaces, and automations). Enterprise customers can pick the underlying model, including OpenAI's GPT family and Anthropic's Claude through Amazon Bedrock. Airtable became a launch partner for ChatGPT's business app directory, so teams can update Airtable bases without leaving ChatGPT.
Microsoft announced Power Apps Copilot in March 2023 at its "Future of Work" event. The pitch was that a user could describe an app, flow, or bot in plain language and Copilot would build it. Copilot in Power Apps reached broader availability over 2024, with new "app skills" (data entry, exploration, visualization, summarization) reaching general availability in 2024 and 2025.
Visual Copilot converts Figma designs into production-ready code in React, Vue, Svelte, Angular, Qwik, Solid, or HTML. The pipeline uses a specialized model (trained on more than 2 million data points) plus the open-source Mitosis compiler, then a final framework-specific pass. Component mapping lets the tool reuse a user's existing component library instead of generating fresh markup.
Apple announced two AI features for Xcode 16 at WWDC 2024:
Swift Assist had a rough rollout. Apple originally promised it for later in 2024 but it slipped past the WWDC 2025 keynote, drawing public criticism. MacRumors and others reported on its absence in March 2025. Predictive code completion shipped on time with the Xcode 16 release in September 2024.
At WWDC 2025, Apple opened the Foundation Models framework to third-party developers, exposing direct API access to the on-device model that powers Apple Intelligence. The framework gives any iOS app access to text summarization, content generation, and intelligent suggestions without sending data to a server.
Google introduced Gemini Pro into Android Studio in April 2024 and expanded the integration at Google I/O 2024 in May. Gemini in Android Studio writes code from custom prompts, explains code, generates unit tests, helps with build errors, and writes release notes. Gemini 1.5 Pro arrived in Android Studio later in 2024, bringing a large context window and multimodal input.
Gemini in Android Studio is now part of Gemini Code Assist, Google's broader developer offering that also covers VS Code, IntelliJ, Colab, Firebase, and Cloud Workstations.
At Google I/O 2024, Google released the new Vertex AI for Firebase SDKs for Dart, Kotlin, Swift, and JavaScript, so a Flutter developer can call Gemini directly from a cross-platform codebase. Flutter 3.22 launched alongside Dart 3.4 with stable WebAssembly compilation for Flutter web.
Google integrated Gemini models into the Play Store starting in 2024. AI-generated review summaries appear under "Users are saying" on each app listing, along with filter chips that surface reviews about specific features (UI, app stability, account issues). Google also generates app highlights and FAQs.
Research on AI-summarized reviews shows two side effects. A University of Hawaii study on travel platforms found that AI summaries reduce both the number and the length of later user reviews, and they correlate with higher average ratings. The interpretation is that users feel less need to write a long review when the summary already captures common themes.
Apple added AI-powered review summaries to the App Store with iOS 18.4 in March 2025, mirroring Google Play's feature.
The App Store also introduced new privacy and AI transparency rules:
PrivacyInfo.xcprivacy file, announced at WWDC 2023 and enforced from May 1, 2024, makes developers declare exactly what data their app collects and which sensitive APIs it uses.The combination of vibe-coding tools and stricter privacy review created tension. Apple's App Review team became overloaded in 2025 as submission volume hit 557,000 new apps, and the company began pushing back on AI-generated apps that failed transparency requirements.
A text-to-app builder is only useful if the backend can keep up. Several platforms became default destinations for AI-generated apps in 2024 and 2025.
Supabase is an open-source Firebase alternative built on PostgreSQL. Supabase Studio added an AI Assistant in 2024, accessible with cmd+i. The Assistant v2 release in December 2024 added schema design, data queries and charting, error debugging, Postgres RLS policy creation, function and trigger management, and SQL-to-supabase-js conversion. Only the schema structure is sent to the underlying model, not the data itself. Supabase is the default backend for Lovable and a common choice for Bolt.new and Codev.
Convex is a TypeScript-first backend platform with cloud functions, a database, file storage, scheduling, vector search, and real-time updates in one package. Its Agent component stores conversation history per user thread, supports tool calling with reactive queries, and ranks messages with hybrid text + vector search. Convex's pitch is that agentic workflows can run in the background while the UI stays reactive.
The Vercel AI SDK is a TypeScript toolkit for building AI-powered apps on React, Next.js, Vue, Svelte, and Node. It standardizes provider integration so an app can switch between OpenAI, Anthropic, Google, Cohere, xAI, and Groq with one line of code. AI SDK 4.0, released in November 2024, added PDF support, computer use integration, Grok support, prompt caching for OpenAI, and a feature that detects incomplete responses and continues generation in multiple steps.
Cloudflare announced Workers AI on September 27, 2023 at its Birthday Week. The pitch was "serverless GPU inference at the edge". Models run on GPUs in Cloudflare data centers near the user; developers pay per request. The launch model list included Meta's Llama 2 7B and OpenAI's Whisper for speech recognition. Cloudflare planned roughly 100 GPU sites by end of 2023 and near-global coverage by end of 2024. Workers AI integrates with Cloudflare Workers and Pages and is also reachable from any platform via REST API.
CodeSandbox is a browser IDE with around 4.5 million monthly users. It rolled out an AI coding assistant called Boxy, then deprecated it in July 2024 in favor of a free integration with Codeium. The CodeSandbox SDK lets developers programmatically spin up sandboxed microVMs in about three seconds, which became important as AI agents needed disposable environments to run untrusted generated code. Together AI acquired CodeSandbox in December 2024 to power its code interpreter.
Marketplaces for AI-generated apps emerged alongside the tools that built them.
OpenAI launched the GPT Store on January 10, 2024 as part of a new ChatGPT Team plan. The store catalogs custom GPTs across categories including lifestyle, writing, research, programming, and education. GPTs are built without code through the GPT Builder tool. Launch partners included AllTrails (trail recommender), Khan Academy (code tutor), and Canva (content designer). Access requires a paid ChatGPT plan (Plus, Team, or Enterprise). OpenAI launched a revenue program for U.S. builders in Q1 2024 based on user engagement.
Claude Artifacts launched in preview in June 2024 alongside Claude 3.5 Sonnet and reached general availability on August 27, 2024 across Free, Pro, and Team plans. An Artifact is a self-contained piece of generated content (code, SVG, HTML page, document, diagram, dashboard) that appears in a side panel next to the chat. Tens of millions of Artifacts had been created by the August general-availability announcement. In June 2025 Anthropic upgraded Artifacts so that they can call the Claude API directly from inside the Artifact, allowing users to build and share fully AI-powered apps.
Replit's combination of cloud IDE plus mobile app publishing has become an informal marketplace. Mobile Apps on Replit, launched in February 2025, lets users go from prompt to App Store with three clicks. By the end of 2025, Replit had more than 40 million users.
Wix's acquisition of Base44 in June 2025 effectively made Wix's app and site marketplace a vibe-coding destination. Wix is moving Base44 technology into the main Wix builder so that a user can describe an app on the same platform that hosts and sells it.
Testing and asset generation followed the same automation curve.
The Snyk and Retool security teams independently published reports in 2025 showing that vibe-coded apps ship with predictable vulnerabilities. Escape.tech analyzed 5,600 publicly deployed vibe-coded apps and found more than 2,000 vulnerabilities and 400 exposed secrets. A December 2025 study of five major AI coding tools across 15 apps found 69 vulnerabilities, with patterns consistent enough to suggest a structural problem rather than user error.
The common failure modes include:
Bolt.new and Lovable both added warnings and templates intended to nudge users toward safer defaults, but the underlying issue is that the model writes code that compiles and ships, not necessarily code that is safe.
A March 2026 TechCrunch report cited a study that compared AI-powered apps with non-AI apps in the same categories. AI apps showed cancellation rates 30% faster than non-AI apps at the median, and refund rates of 15.6% versus 12.5%. The interpretation is that AI features alone do not drive long-term retention if the underlying app is not solving a real user problem.
Apple received 557,000 new app submissions in 2025. Roughly 70% of App Store users discover apps through search, and 90% never scroll past the tenth result. Real developers are losing visibility to AI-generated apps that get indexed first. App review teams have stretched review wait times from days to weeks. Apple has begun blocking updates for some AI vibe-coded apps that fail privacy transparency rules.
Inference is the recurring operational cost every time the model generates a response. By 2025, roughly 80% of AI app budgets went to inference and 20% to training and infrastructure. Inference costs for a GPT-3.5-class system dropped more than 280-fold between November 2022 and October 2024, but total AI spend in production went up over the same period because usage grew faster than unit prices fell.
At small scale, an AI app costs roughly $0.02 per active user per month for chat-style workloads (10 requests, 1,000 tokens each, at $0.002 per token). Agentic workflows cost five to twenty-five times more than chat. Single agentic tasks run $0.10 to $0.50 each, which scales to $150,000 to $750,000 per month at 10,000 users. Between 500 and 2,000 concurrent users, many teams move from cloud APIs to self-hosted GPUs, and costs jump three to five times overnight. By 5,000 to 10,000 users, margins on many AI features collapse without aggressive optimization (model routing, prompt caching, response caching, on-device inference where possible).
Replit switched from $0.25 per checkpoint to effort-based pricing in mid-2025. The new pricing rolled out to new users on June 18 and to all users by July 2. Some long-time users complained on InfoWorld and on Replit's own forums that the new model was harder to predict and made hobby projects unaffordable. Replit defended the change as more accurate, since complex requests now bill as a single meaningful checkpoint instead of several intermediate ones.
Marketing for text-to-app tools often implies that any user can ship a real product without a tutorial. The reality, as Snyk, Retool, and Glide all observed in 2025, is that the tools can produce working frontends but produce backends with the security and operational defaults of a hackathon demo. Some critics argue the tools have actually raised the bar for production-grade development, because the easy part (UI scaffolding) is now free while the hard parts (data security, authentication, observability, billing) still require traditional engineering judgment.