Figma to Production Code: How We Turn Designs Into Live React Apps

TL;DR
AI can generate about 70% of a React app from a Figma file. The remaining 30% — state handling, error and empty states, accessibility, and performance — is where production apps are won or lost. This is the six-step Figma to production code workflow we use on client projects, including design tokens and Code Connect.
Figma to Production Code: How We Turn Designs Into Live React Apps
Turning Figma to production code is still one of the most painful steps in building a product. Figma's own research found that 92% of designers and developers agree the handoff needs improvement. That number has barely moved in years, even as the tools got better.
The reason is simple. A design file and a codebase are two different things. A Figma frame describes how something looks at one size. A React app has to handle every size, every state, every error, and every user. Closing that gap is engineering work, not export work.
At Craxinno, we have shipped this handoff across 210+ projects. This post explains exactly how we do it: the preparation that makes designs convertible, the tooling we use, the part AI handles well, and the part that still needs a developer. If you are evaluating how to get your designs into a live app, this is the honest version.
Why most Figma to React conversions fail
Before the process, it helps to know where teams go wrong. Three failures show up again and again.
The design was never built for code. A designer creates a beautiful screen using absolute positioning and one-off colors. It looks perfect at one width. In code, it breaks on mobile, because nothing tells the layout how to flex. The fix belongs in Figma, not in the codebase.
Every value is hard-coded. The design uses 47 slightly different hex values and a dozen padding numbers. The developer copies each one by hand. Six months later the brand changes, and someone has to hunt down every instance. This is the single biggest source of long-term drift.
The export was treated as the finish line. A tool generates a React component that looks right in a screenshot. Nobody checks keyboard access, loading states, or what happens when the API returns nothing. The app looks done and behaves broken.
Our Figma to production code workflow, step by step
Here is the process we run on client work. It is deliberately front-loaded, because the preparation is what makes the rest fast.
Step 1: Audit the design file before writing any code
We open the Figma file and check three things. Is auto layout used on every frame that needs to be responsive? Are components named the way they will exist in code? Are colors, type, and spacing defined as styles and variables rather than one-off values?
If the answer is no, we fix that first. This step feels like a delay. It is the opposite. A file that is properly structured converts in a fraction of the time, and the code that comes out is maintainable.
Step 2: Extract design tokens as the single source of truth
We pull colors, spacing, typography, and radii out of Figma as design tokens, then map them into the codebase as CSS variables or Tailwind theme keys.
This is the highest-leverage step in the entire workflow. Once every value in the code traces back to a named token, a rebrand becomes a single change in one file instead of a search across fifty components. It also removes ambiguity between teams. When a designer says primary-600 with spacing-4, the developer knows exactly what that means. No screenshots, no back-and-forth.
Step 3: Map Figma components to React components one-to-one
A button in Figma with size and variant properties should become a React Button component with size and variant props. Same name, same structure, same options.
We use Figma Dev Mode and Code Connect to make that mapping explicit. The payoff is that the design system and the codebase stop drifting apart. A change in one has an obvious home in the other.
Step 4: Generate the first pass with AI, then engineer the rest
This is where we are direct about what AI does and does not do. Modern tooling, including Figma's MCP server and AI coding assistants, can read component properties, tokens, and auto layout directly from the file and produce a solid first draft. Claude and Claude Code are part of our daily workflow here, and they genuinely compress the tedious part.
But the first pass is roughly 70% of the job. It gets structure and styling close. It does not reliably handle state management, error boundaries, loading and empty states, data fetching, or accessibility. Research cited across the industry found automated accessibility tools catch only about 30% of WCAG issues in generated output. Engineering cleanup is not optional. It is the work.
Step 5: Build the states the design never showed
Every design shows the happy path. Production needs the rest. For each screen we build the loading state, the empty state, the error state, and the partial-data state.
This is the step that separates a demo from a product, and it is the one most conversion tools skip entirely because a Figma file simply does not contain that information.
Step 6: Test responsiveness, accessibility, and performance
We check the app at real breakpoints, not just the two the design showed. We verify keyboard navigation, focus order, colour contrast, and screen reader labels. We measure bundle size and load performance.
Then we ship it, usually on Next.js and Vercel, with the component library documented so the next developer inherits something clean.
The stack we use for Figma to React work
Our production stack is deliberately standard, because standard means maintainable by whoever comes next.
For the design side: Figma with Dev Mode, Code Connect for component mapping, and design tokens exported as the source of truth.
For the code side: React and Next.js with TypeScript, Tailwind for styling driven by those tokens, and a documented component library.
For the AI-assisted layer: Claude and Claude Code for first-pass generation and refactoring, always followed by human review.
There is no exotic lock-in here. If you take the codebase elsewhere, any competent React team can pick it up. That is intentional.
When to use a conversion tool, and when not to
Honest guidance, since this comes up on nearly every project.
Use a one-shot conversion tool when you need a marketing page, a landing page, an internal tool, or a throwaway prototype. Tools like Anima, Locofy, and Visual Copilot get you to a working build quickly, and the output is good enough for something you will not maintain for years.
Do not architect a product on them. For an application you will grow, hire on, and maintain, generated output tends to produce code that works but is unpleasant to extend. The cost shows up later, in every feature that takes longer than it should.
The rule we use: if the code will live longer than six months or be touched by more than two developers, engineer it properly.
What good Figma to code handoff actually delivers
When the workflow above is done well, the outcomes are concrete.
Rebrands take hours instead of weeks, because tokens live in one place. New screens ship faster, because the component library already exists. Visual bugs get rarer, because values are named rather than copied. And designers and developers stop arguing about padding, because they share a vocabulary.
That is the real return on the handoff. Not pixel accuracy on day one, but speed on day ninety.
Work with a team that ships design to production
If you have Figma designs and need them live as a fast, accessible, maintainable React app, the Craxinno team does this work daily. See recent builds in the Craxinno portfolio, view our full capabilities on the services page, or email hello@craxinno.com.
Frequently Asked Questions
Can AI convert Figma designs to production React code?+
AI can generate a strong first pass, typically around 70% of the work. Tools using Figma's MCP server read component properties, design tokens, and auto layout directly from the file. But generated code rarely handles state management, error and empty states, data fetching, or accessibility properly. Automated accessibility tools catch only about 30% of WCAG issues, so engineering review is still required.
What are design tokens, and why do they matter for Figma to code?+
Design tokens are named values for colors, spacing, typography, and radii, shared between the design file and the codebase. They matter because they remove hard-coded values. When every value traces back to a token, a rebrand is one change in one file instead of a search across dozens of components.
How do you prepare a Figma file for code conversion?+
Use auto layout on every frame that needs to be responsive, since absolute positioning converts to fixed pixels and breaks on mobile. Name components the way they will exist in code, so a frame becomes a readable component. Define colors, type, and spacing as styles and variables rather than one-off values.
Should I use a Figma to React conversion tool?+
Use one for marketing pages, landing pages, internal tools, and prototypes, where speed matters more than long-term maintainability. Avoid architecting a real product on generated output. A good rule: if the code will live longer than six months or be touched by more than two developers, engineer it properly.
How long does Figma to production code take?+
It depends mostly on how well the design file is prepared. A well-structured file with auto layout, named components, and defined styles converts far faster than an unstructured one. A single marketing page can ship in days. A full application with a component library, states, and accessibility work typically takes weeks.
Need a site that performs?
Fast, accessible web apps in React and Next.js — built to rank and built to scale.
Start a projectKeep ReadingMore case studies like this
Engineering retros, product launches, and brand systems from our studio — updated monthly.
All case studiesTechnology Used
Tags & Keywords
Continue with Blogs.
View all blogs
AI Agent DevelopmentAI Agent Development Services in India: Pricing, Timelines & Engagement Models
Short answer: In 2026, AI agent development services in India typically cost $4,000–$8,000 for a single-task proof of concept, $12,000–$24,000 for a production agent with real integrations, and $30,000–$60,000+ for a multi-agent system. Timelines run 3 to 24 weeks. Indian teams typically bill $15–$70 an hour, against $100–$200 for US agencies. Those numbers are worth very little on their own, though. Two quotes of $15,000 can buy completely different things: one buys a chatbot with a prompt, the other buys an agent that reads your CRM, takes actions, logs every decision and fails safely. This guide breaks down what you actually pay for when you hire for AI agent development services in India , how long each stage takes, and which engagement model fits your situation. We're Craxinno , an AI-first design and development studio in Jaipur. We bill AI agent work at $25/hour, so the figures below come from how we scope real projects, plus published ranges from other Indian and global vendors for comparison. AI agent development pricing in India at a glance Every serious quote is hours multiplied by a rate, plus the monthly cost of running the thing. Here's how that maths works out at an Indian studio rate of $25/hour (roughly ₹95 to the dollar at the time of writing, so check the rate on the day you sign). Agent type What it does Typical effort Build cost at $25/hr Timeline Proof of concept One task, one or two tools, basic memory. Answers from your docs or drafts replies for a human to approve. 160–320 hours $4,000–$8,000 (₹3.8–7.6 lakh) 3–5 weeks Production agent One complete workflow, 2–4 system integrations, guardrails, logging, evaluation suite. 480–960 hours $12,000–$24,000 (₹11.4–22.8 lakh) 8–12 weeks Multi-agent system Several specialist agents with an orchestrator, many integrations, human approval steps, audit trails. 1,200–2,400+ hours $30,000–$60,000+ (₹28.5–57 lakh+) 14–24 weeks For context, global agencies quote far higher for similar scopes. Master of Code puts most agent builds between $15K and $350K, and our own breakdown of what it costs to build an AI agent in 2026 covers the Western agency ranges in detail. The gap is almost entirely labour. Which brings up the part most "India is cheaper" articles skip. Why AI agents cost less in India (and the part that doesn't) India's advantage is engineering hours. Second Talent's 2026 rate benchmark puts a mid-level AI agent developer in India at $28 to $72 an hour, against $119 to $185 in the US. Studios with lower overheads (Craxinno included) come in under that Indian range. On a 600-hour production agent, $25/hour versus $150/hour is the difference between a $15,000 invoice and a $90,000 one. But your OpenAI, Anthropic or Gemini bill is priced in dollars, and it's the same whether your engineers sit in Jaipur or Boston. Same for your AWS account, your vector database and your observability tools. So the saving applies to the build . It doesn't apply to the run . A team that tells you an Indian-built agent will also be cheaper to operate is either planning to use smaller models (fine, if they say so) or hasn't done the token maths. We wrote up the ways to cut LLM API costs separately, because that's where long-term savings actually come from. What actually moves the price Agent quotes swing wildly because "AI agent" covers everything from a FAQ bot to software that books freight. When we scope a project, four things decide where it lands. 1. How many systems the agent touches Each integration (CRM, ERP, helpdesk, payment gateway, WhatsApp Business API, a legacy SQL database nobody documented) adds roughly a week of work: authentication, error handling, rate limits, test data. A clean REST API with good docs is fast. A SOAP endpoint from 2011 is not. 2. Whether it reads or acts An agent that only reads and answers is low-risk. One that writes to your systems (issues refunds, updates orders, sends emails) needs permission scopes, confirmation steps, rollback logic and far more testing. Write access is usually the single biggest jump in cost. 3. How ready your data is RAG agents live or die on the documents behind them. If your knowledge base is 400 PDFs with three conflicting versions of the refund policy, someone has to clean that up first. Budget a week or two for it, or the agent will confidently quote the wrong policy. 4. Evaluation and guardrails This is the line item cheap quotes quietly drop. A production agent needs a test set of real questions, scored answers, regression checks every time a prompt or model changes, and monitoring once it's live. It's maybe 15–20% of the build effort. Skip it and you find out about failures from your customers. That last point matters more than it sounds. Gartner predicts that over 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value and weak risk controls. Most of those won't fail because the model was bad. They'll fail because nobody measured whether the agent was doing its job. AI agent development timeline: week by week Here's how a typical 10-week production agent breaks down at our studio. A proof of concept compresses the middle; a multi-agent system stretches it. Phase Duration What happens What you get Discovery and scoping 1 week Pick one process, define success metrics, map integrations, check data quality Scope doc with an hour estimate per item Prototype 1–2 weeks Core agent loop working on real data, no polish A demo you can poke at Build and integrate 4–6 weeks Tools, integrations, memory, admin views, guardrails Bi-weekly demos on a staging environment Evaluation and hardening 2 weeks Test sets, failure analysis, prompt and model tuning, load checks Accuracy report against the agreed metrics Launch and hypercare 1–2 weeks Production deploy, monitoring, fast fixes on live traffic A live agent plus a runbook The prototype week is where most projects either earn their budget or get cut. We'd rather you discover in week three that your use case doesn't need an agent (sometimes a well-built workflow automation is enough; see our explainer on AI agents vs chatbots ) than in week ten. Be wary of "live in 5 days" promises. That's usually a no-code template with your logo on it. It can be the right call for a simple FAQ bot. It isn't a custom agent. Engagement models for AI agent development services in India Indian studios generally offer four or five ways to work together. The right one depends on how well you can define the scope today, not on which looks cheapest. Model How billing works Best for Who carries the risk Fixed price One agreed price, paid in milestones Clearly scoped agents where integrations are known Mostly the vendor, so expect a buffer in the price Time and materials Hourly rate, billed on logged hours Exploratory builds and evolving requirements Mostly you, offset by full visibility Dedicated team Monthly fee per engineer, full-time on your project Ongoing AI roadmaps, multiple agents, product companies Shared White-label Studio builds under your agency's brand Agencies reselling AI work to their own clients Shared, governed by the subcontract Support retainer Monthly block of hours after launch Monitoring, prompt updates, model upgrades, new flows Shared Fixed price Good when you can describe the agent's job in a paragraph and list every system it touches. You know the number up front. The trade-off is that changes cost a change request, and any honest vendor builds in a contingency buffer because agent behaviour is harder to predict than a CRUD app. Time and materials Most first AI agent projects should start here, at least for discovery and the prototype. You pay for hours worked, you see timesheets, and you can redirect the team the moment a test result surprises you. At Craxinno this is $25/hour with hours shown in each bi-weekly report. Dedicated team Once you have one agent in production and a list of five more, a dedicated engineer or pod usually works out cheaper than a series of fixed bids. Our dedicated development team cost guide walks through the monthly maths. Longer commitments are also where Indian studios have room to go below their headline rate; we offer $15–$20/hour for longer, steadier engagements. White-label If you run a marketing, design or software agency abroad and your clients keep asking for AI agents, white-labelling lets you sell the work without hiring an AI team. The studio stays invisible. Your margin is the gap between your client rate and the studio's rate. Support retainer Agents need maintenance more than websites do. Model providers deprecate versions, prices change, and your business rules drift. A small monthly retainer covers monitoring and keeps the agent honest. Our recommendation: start with a time-and-materials discovery and prototype (3–4 weeks), then switch to fixed price for the production build once the unknowns are gone. You get flexibility while you're still learning, and price certainty once you aren't. Running costs after launch The build is a one-time cost. These are not. Budget for them before you sign, using ranges from our AI agent pricing guide and the cost-to-build breakdown linked above: Model usage: around $100–$500/month for a light agent, and $2,000–$10,000+ for a heavy autonomous one Cloud hosting: $200 to several thousand a month depending on traffic Vector database (if you use RAG): a few hundred dollars and up Maintenance: plan on 15–30% of the build cost per year A sensible rule: if a vendor can't estimate your monthly running cost before you commit, they haven't designed the agent yet. How to vet an AI agent development company in India Thousands of Indian firms now sell AI agent development services, and plenty of them added "AI" to their homepage last year. A handful of questions separate the builders from the rebranders: "Show me an agent you shipped in the last six months, running." A recording is fine. Slides aren't. "How will you measure whether it works?" You want to hear about test sets and accuracy targets, not "we'll do UAT." "What will this cost per month to run?" Covered above. No answer, no deal. "Who owns the code, prompts and evaluation data?" It should be you, in writing. We've written longer checklists on how to choose an AI agent development company and hiring AI developers in India if you're comparing shortlists. Time zones come up a lot too. Jaipur is 9.5 to 10.5 hours ahead of New York depending on daylight saving, which sounds awkward. In practice it means a short overlap call in your morning and a day's work waiting for you the next one. How Craxinno prices and runs AI agent projects We keep this simple on purpose: Rate: $25/hour for AI agent work, with $15–$20/hour available on longer commitments Process: one week of scoping, then 8–14 weeks of design and build for most production agents, with a demo every two weeks ( see our work process ) Transparency: hour estimates per feature up front, and a running-cost estimate before you commit Ownership: code, prompts and evaluation sets are yours Stack: Python (FastAPI, Flask, Django), RAG pipelines, agentic workflows, cloud deployment on AWS and RunPod We design and ship, so the admin screens and user-facing parts of your agent come from the same team that builds the backend. No hand-off to a separate design agency. If you have a process in mind, send us a short description . We'll come back with an hour estimate, a timeline and a monthly running-cost figure, so you can compare us against anyone else on the same terms.
FlutterFlutter vs. React Native vs. Kotlin Multiplatform (2026)
Flutter vs. React Native vs. Kotlin Multiplatform (2026) Pick a cross-platform framework at the end of 2026 and you aren't choosing between three flavours of the same idea. Flutter draws every pixel with its own engine. React Native renders real native components from JavaScript. Kotlin Multiplatform shares Kotlin code and lets you decide how much of the UI to share, if any. This guide compares Flutter vs React Native vs Kotlin Multiplatform on the versions shipping right now (Flutter 3.47, React Native 0.88, Kotlin 2.4 with Compose Multiplatform 1.11), not the 2023 benchmarks most articles still recycle. We build web and mobile products at Craxinno, so the verdict is opinionated. You'll know which one to pick before you reach the table. Quick verdict: which framework should you pick in 2026? Choose Flutter if: your design is the product: custom UI, heavy animation, a brand look that has to match pixel for pixel on iOS and Android you want one team and one codebase for mobile plus desktop, TV or in-car screens you're happy to hire Dart developers, or train them (Dart takes a competent developer a couple of weeks) you're building a white-label app that gets re-themed for many brands Choose React Native if: your team already writes React and TypeScript for the web you want the biggest hiring pool and over-the-air fixes through Expo your app is mostly content, commerce, dashboards or forms rather than custom graphics Choose Kotlin Multiplatform if: you already have a native Android app (or two native apps) and you're tired of writing the same business logic twice iOS has to feel exactly like iOS, with SwiftUI on the screens that matter the hard part of your app is logic: offline sync, payments, encryption, on-device AI you'd rather adopt cross-platform one module at a time than rewrite Our default: for a new MVP from a web-first team, React Native with Expo. For a design-led consumer app, Flutter. For an existing native app that's getting expensive to maintain twice, Kotlin Multiplatform. Not sure which bucket you fall into? Send us your feature list and we'll give you a straight answer, even if the answer is "go native." Flutter vs React Native vs Kotlin Multiplatform at a glance Criteria Flutter React Native Kotlin Multiplatform Backed by Google Meta, with Expo and Microsoft JetBrains, officially supported by Google for Android Language Dart JavaScript / TypeScript Kotlin (plus Swift if iOS UI stays native) Latest stable (Nov 2026) 3.47 with Dart 3.13 (Aug 2026) 0.88 (Oct 2026) Kotlin 2.4, Compose Multiplatform 1.11 How the UI renders Own engine (Impeller) draws every pixel Real native views through Fabric Fully native UI, or shared Compose Multiplatform UI Code sharing Highest by default : UI and logic Most of the app: UI and logic Logic by default, UI optional Native iOS feel Imitated (Cupertino widgets) Native components Native by definition with SwiftUI Hiring pool Medium Largest Smaller, but every Android developer already writes Kotlin Over-the-air updates Third-party (Shorebird) Yes (Expo EAS Update and others) No, shared code is compiled Beyond mobile Web, desktop, embedded Windows, macOS, web, Meta Quest Desktop (stable), web (beta) Best for Custom, animation-heavy apps React teams, content and commerce apps Existing native apps, logic-heavy apps What changed in 2026 (and why older comparisons are wrong) Most "Flutter vs React Native" benchmarks you'll find were run on rendering pipelines neither framework uses any more. Four changes matter. Flutter finished its move to Impeller. Flutter 3.44, released at Google I/O in May 2026, removed the old Skia backend on Android 10 and above. Then Flutter 3.47 (August 12, 2026) made Impeller the default on Windows, macOS and Linux too. Impeller precompiles its shaders, which removes the cause of the first-run animation stutter people used to blame on Flutter. React Native dropped the bridge for good. Version 0.82 (October 2025) was the first release that runs only on the New Architecture. Version 0.84 made Hermes V1 the default JavaScript engine. And React Native 0.88 moved event handling to the web-standard EventTarget API. It was also the third release in a row with no user-facing breaking changes. If you remember what React Native upgrades felt like in 2021, that last point is the headline. Compose Multiplatform for iOS went stable. JetBrains shipped Compose Multiplatform 1.8.0 in May 2025 with iOS marked stable and production-ready. Version 1.11 followed in May 2026. Swift Export reached Alpha. In Kotlin 2.4, Kotlin suspend functions export to Swift as async functions and Kotlin Flows export as AsyncSequence . iOS developers calling shared Kotlin code no longer have to go through an Objective-C header layer for everything. It's still Alpha, so treat it as "coming" rather than "done." We used to describe KMP to clients as "a shared library, not a framework." Since the Compose Multiplatform iOS release, that description is simply out of date. How we compared them We scored each framework on the questions clients actually ask us before a build: How fast will it feel? How much code do we write once? Who can we hire? What does it cost to keep running in year three? Version facts come from each project's official release notes and documentation, linked throughout. We didn't run fresh synthetic benchmarks for this article. Independent 2026 benchmarks disagree with each other depending on the device and the test app, and a single chart would give you false confidence. Where a number appears below, it's linked to its source. The opinions come from shipping React, React Native and web products at Craxinno . Head-to-head comparison Architecture: how each one puts pixels on screen Flutter ships its own rendering engine inside your app. A Flutter button isn't an iOS button; it's a drawing of one. That's why a Flutter app looks identical on a three-year-old Android phone and a new iPhone, and why it can run on surfaces with no native UI toolkit at all, like Toyota's 2026 RAV4 dashboard. React Native runs your JavaScript on Hermes and talks to native code directly through JSI. When you write <View> , iOS gets a real UIView. The New Architecture (Fabric for rendering, TurboModules for native calls) is now the only architecture. Kotlin Multiplatform compiles shared Kotlin to JVM bytecode for Android and to native machine code for iOS through Kotlin/Native. The UI is your choice: Jetpack Compose and SwiftUI on each side, or one shared Compose Multiplatform UI. Most production teams we talk to mix both, sharing utility screens like settings and forms while keeping "hero" screens native. Performance: is Flutter faster than React Native in 2026? Verdict: Flutter wins on animation consistency, KMP wins on heavy logic, and React Native is close enough for most business apps. Picture a trading app's portfolio screen. Four hundred rows, a live chart, prices ticking every second. Flutter handles that kind of screen well because nothing crosses a boundary: Dart builds the frame and Impeller draws it. React Native can match it today, but it asks more of your team (FlashList instead of FlatList, memoisation, animations on the UI thread with Reanimated). Now picture the same app syncing 50,000 transactions offline and reconciling them. That's where Kotlin Multiplatform pulls ahead, because the shared logic is compiled native code, not a script running in an engine. Quizlet moved its grading logic to KMP and reported a 25% faster grading algorithm on iOS compared to JavaScript . And React Native keeps closing the gap release by release. Hermes V1 became the default in 0.84, and 0.88.1 switched iOS release builds from size-optimised ( -Os ) to speed-optimised ( -O2 ) compilation to match Android. For a typical commerce, booking or SaaS companion app, your users won't feel the difference between a well-built React Native app and a well-built Flutter app. They will feel a badly built one in either. Code sharing: how much do you actually write once? Verdict: Flutter shares the most by default. KMP shares exactly what you choose. With Flutter, close to the whole app is shared. You drop to platform code (Swift, Kotlin) only for platform channels: a payment SDK, a Bluetooth device, a widget extension. React Native is similar, with a slightly bigger native slice because it uses real platform components and more third-party native modules. KMP flips the default. Out of the box you share the data layer, networking, validation, business rules and state. Add Compose Multiplatform and you can share UI too; JetBrains lists Respawn Pro as sharing 96% of its code between iOS and Android. The Kotlin Multiplatform Survey 2025 found around 80% of KMP plugin users already use Compose Multiplatform, so "logic only" is no longer the typical setup. Native access and how "iOS" the app feels Verdict: KMP with SwiftUI feels the most native. React Native comes second. Flutter needs the most care. Flutter's Cupertino widgets are a good imitation, but they are an imitation. When Apple changes the look of system controls, Flutter apps wait for the framework to catch up. Flutter 3.44 and 3.47 moved Material and Cupertino into standalone packages ( material_ui and cupertino_ui 1.0), so design updates can now ship outside the quarterly Flutter release. It helps. It doesn't change the underlying model. Whichever framework you pick, you'll still touch native code. When we built the React Native wrapper for Eddyter , our AI rich-text editor, the JavaScript side was the straightforward part. The fix that mattered was on the Android side: speech-to-text needed the MODIFY_AUDIO_SETTINGS permission. That's the reality of cross-platform work. Budget for someone on the team who is comfortable opening Xcode and Android Studio. Developer experience and tooling Verdict: Flutter has the most polished single toolchain. React Native with Expo is the fastest way from zero to a store build. KMP is improving fastest. Flutter's stateful hot reload, DevTools and now-stable Widget Previews (in 3.47) make UI work quick. Everything lives in one SDK, so there's less to wire together. React Native's experience depends heavily on Expo. With Expo you get managed builds, updates and native configuration without touching Xcode for weeks. React Native's own DevTools now includes a real network inspector. Without Expo, you're assembling more of the toolchain yourself. KMP lives in Android Studio and IntelliJ IDEA with the Kotlin Multiplatform plugin, which gives you Compose hot reload, shared UI previews, and navigation and debugging across Kotlin and Swift code. The weak spot is still the iOS build loop through Kotlin/Native, which is slower than an Android build and something JetBrains keeps working on. Hiring and team cost Verdict: React Native is the easiest to staff. KMP is the easiest to staff if you already have Android developers. React Native draws from the enormous pool of React web developers. In the UK, Foresight Mobile counts roughly 3.4 times more available React Native developers than Flutter developers , with Flutter developers paid around 10 to 15% more. A 2026 job-listing analysis by Quash showed the same React Native lean in Bengaluru. Flutter developers are fewer, but Dart is easy to learn and Flutter developers tend to be strong on UI. Kotlin Multiplatform's hiring story is different: you're not hiring "KMP developers," you're extending the Android developers you already have. JetBrains reports that over 20,000 companies use KMP, including Google, Duolingo, Booking.com, McDonald's and Philips. The catch: if your iOS UI stays in SwiftUI, you still need an iOS developer. If you're hiring offshore, our guide to what a dedicated development team costs breaks down the team shapes in more detail. Web, desktop and beyond mobile Verdict: Flutter reaches the most screens. Flutter targets mobile, web, Windows, macOS and Linux from one codebase, with Canonical now driving the desktop work. It's also moving into embedded screens: LG is bringing Flutter to webOS smart TVs and Toyota runs it in the 2026 RAV4. React Native covers Windows and macOS through Microsoft's forks (React Native Windows 0.85 shipped in October 2026), the web through React Native Web, and Meta Quest headsets. In practice, most teams with a web product keep a separate React or Next.js site and share TypeScript types, validation and API clients with the app. Compose Multiplatform is stable on Android, iOS and desktop, with web (Kotlin/Wasm) still in beta. Good for internal desktop tools; not yet what we'd pick for a public website. Over-the-air updates Verdict: React Native wins clearly. Found a bug in checkout on a Friday night? With React Native and Expo EAS Update you can push a fixed JavaScript bundle to users without waiting for App Store review (within Apple's and Google's rules: fixes and small changes, not new features that change what the app is). Flutter has a third-party option in Shorebird. Kotlin Multiplatform has no equivalent for compiled code, so every fix ships through store review. Long-term risk All three are safe bets in 2026, for different reasons. Flutter keeps a steady four-stable-releases-a-year cadence with a published roadmap. React Native is used in Meta's own products and is co-maintained with Expo and Microsoft. KMP is JetBrains' strategic bet and Google officially supports it for sharing business logic on Android. The bigger risk is the one teams create themselves: picking a framework their team can't hire for in two years. Which framework fits your situation? You're a seed-stage startup with a Next.js web app. Pick React Native with Expo. Your web developers can contribute from week one, and you can share TypeScript types and validation with the web app. We cover the rest of that stack in our 2026 SaaS tech stack picks . You're launching a consumer app where the brand look is the selling point. Pick Flutter . Custom transitions, illustrated onboarding and a pixel-identical design on both stores are where Flutter earns its keep. You're a bank, fintech or healthcare company with native apps already in the stores. Pick Kotlin Multiplatform . Move your business rules into a shared module first, keep your SwiftUI and Compose screens, and grow from there. No rewrite, no big-bang migration. You're building one app that gets re-skinned for many clients. Pick Flutter. Theming a single widget tree for ten brands is far simpler than maintaining ten sets of native styles. You're building field-service or warehouse tablets plus a desktop admin tool. Flutter or Compose Multiplatform. Both cover mobile and desktop well; choose by whether your team leans Dart or Kotlin. You need deep platform features on day one (CarPlay, complex widgets, ARKit, background audio). Seriously consider KMP with native UI, or plain native development with Swift and Kotlin. Pros and cons Flutter Pro: identical, smooth UI everywhere, with Impeller now default on every platform Pro: widest reach: mobile, web, desktop and embedded from one codebase Con: iOS look and feel is imitated, so Apple design changes arrive late Con: smaller hiring pool, and Dart skills don't transfer to your web stack React Native Pro: reuses React and TypeScript skills your web team already has Pro: over-the-air updates and the largest library ecosystem through npm and Expo Con: performance on complex screens depends on discipline (list virtualisation, UI-thread animation) Con: third-party native modules are still the most common source of upgrade pain Kotlin Multiplatform Pro: truly native UI and native-speed shared logic Pro: adopt it one module at a time inside existing apps Con: with native UI you still build two interfaces and need iOS skills Con: Swift Export is Alpha and Compose for web is beta; the ecosystem is younger What each option costs to build The hourly rate rarely changes with the framework. At Craxinno we bill the same rate for Flutter, React Native and Kotlin work, so the cost difference comes from hours and team shape, not the stack name. Flutter React Native Kotlin Multiplatform Typical MVP team 1–2 Flutter developers + designer 1–2 React Native developers + designer (web devs can help) Kotlin developer + iOS developer (native UI), or Kotlin developers only (Compose Multiplatform) Where the hours go Custom UI, platform channels for native SDKs Native modules, performance tuning, upgrades Two UIs if native, Kotlin/Native build setup Cheapest when One design across every platform You already have React code and developers You already have an Android app Most expensive when The app must follow iOS design closely The app is graphics- or animation-heavy You're starting from scratch with no Kotlin team For a worked example of how features drive the budget, see how much it costs to build an app like Instagram . Our recommendation Start from your team, not from a benchmark. A React team will ship a better React Native app than a Flutter app, and an Android team will ship a better KMP app than either. When there's no team yet, decide by the product: design-led and multi-screen points to Flutter; content, commerce and fast iteration points to React Native; logic-heavy with uncompromising native iOS points to Kotlin Multiplatform. The one choice we'd avoid in 2026 is picking a framework because a 2022 chart said it was faster. All three are fast now. The question is which one your team can build, hire for and maintain for the next five years. Want a second opinion on your app? Tell us what you're building and Craxinno's team will recommend a stack, a team shape and a realistic timeline. You can also browse our recent projects first.
OpenAI APIHow to Cut Your OpenAI / LLM API Bill (Practical Guide)
Quick answer: To cut your OpenAI or LLM API bill, send each task to the smallest model that passes your tests, put the stable part of every prompt first so it gets cached at up to 90% off, move anything that can wait to the Batch API or Flex (50% off), and cap output tokens. Stacked together, these levers routinely remove most of the bill. Most teams overpay for the same reason. They picked one model during the prototype, wired it into every feature, and never looked again. The invoice grows with traffic, someone in finance asks why, and the answer is "we're using AI." That answer stops working around the third month. This guide walks through the levers we actually reach for when an LLM API bill gets out of hand, in the order we reach for them. Prices quoted are OpenAI's published Standard-tier rates as of November 2026. They change often, so treat the ratios as the lesson and check the official pricing page before you plan a budget around them. Where your LLM API money actually goes Before cutting anything, know what you're paying for. Every request bills three things: input tokens, output tokens, and (on newer models) cache writes. Output is the expensive one. Across OpenAI's current lineup, output tokens cost 5x as much as input tokens, and reasoning tokens are billed as output even though you never see them. Model (Standard tier) Input / 1M Cached input / 1M Output / 1M GPT-6 Astra $10.00 $1.00 $50.00 GPT-6.1 Sol $2.00 $0.10 $10.00 GPT-6.1 Terra $1.00 $0.05 $5.00 GPT-6 Luna $0.10 $0.01 $0.50 GPT-5.4 mini $0.75 $0.075 $4.50 Look at the spread. Astra input costs 100x what Luna input costs. If a feature is running on the top model by default, that one decision probably outweighs every prompt tweak you could make. 1. Measure cost per feature before you cut anything You can't fix a number you only see as a monthly total. Log, for every call: the feature or endpoint, the model, input tokens, cached tokens, output tokens, and the user or tenant. The API already returns most of this in the usage object, including usage.prompt_tokens_details.cached_tokens . Then sort by spend. In almost every audit, one or two features own the majority of the bill. Often it's something nobody thinks of as "the AI feature" (a background summarizer, a retry loop, an agent that re-reads the same document on every step). Start there. 2. Route each task to the smallest model that passes your tests If you want to cut your OpenAI API bill fast, this is the biggest lever and the most ignored one. Classification, extraction, tagging, short summaries, routing decisions and formatting don't need a frontier model. A mini or nano tier handles them, and the price difference is an order of magnitude or more. The trick is to make the decision with data instead of vibes: Pull 100 to 200 real requests for the feature, with the outputs you consider correct. Run them through the cheapest candidate model. Score the results the same way you'd score the expensive model. Move up one tier only if the cheap one fails your bar. For mixed traffic, put a small router in front: the cheap model handles the request, and only escalates to the larger one when it's unsure or the task is flagged as hard. Most requests never escalate. (One honest caveat: a cheaper model that needs three retries isn't cheaper. Count retries in your eval.) 3. Structure prompts so they get cached Prompt caching is the closest thing to free money in the API. When the start of a request exactly matches a recent one, the matching tokens bill at the cached rate, which is 90% to 95% below the normal input price on current OpenAI models. OpenAI applies it automatically, but only if your prompts are built for it. Rules that decide whether you get a cache hit: Stable content first, variable content last. System instructions, tool definitions, schemas and few-shot examples go at the top. The user's message and anything dynamic go at the bottom. No timestamps or request IDs near the top. One changed character early in the prompt breaks the match for everything after it. Move that data into the request's metadata field. Keep tool lists identical. If you need to restrict tools per turn, use allowed_tools in tool_choice instead of changing the tools array. Append, don't edit. Rewriting earlier messages in a conversation throws away the cached prefix. Use prompt_cache_key . It groups related requests onto the same machines. In OpenAI's Prompt Caching 201 cookbook , one coding customer raised their hit rate from 60% to 87% just by adding it. On older models, caching only applies once the prompt reaches 1,024 tokens, so a short prompt never benefits. Here's the part most guides skip. On the GPT-5.6 and GPT-6 families, OpenAI now lists a separate cache-write price: 1.25x the input rate for a 30-minute cache and 2.5x for 12 hours. That changes the math. A prefix you write once and never reuse costs more than not caching. On GPT-6.1 Sol, though, one reuse already wins: $2.50 to write plus $0.10 to read, against $4.00 for sending the same million tokens twice uncached. So caching pays off on any prefix that gets hit at least twice inside the window, which describes nearly every chatbot and agent. 4. Move anything that can wait to the Batch API If the user isn't waiting for the answer, you shouldn't pay real-time prices. The Batch API takes a file of requests, returns results within 24 hours, and bills at 50% of the Standard rate on input, cached input and output alike. Good candidates: Nightly summarization or tagging of tickets, reviews or emails Generating embeddings or metadata for a content backlog Running evals and regression tests on new prompts Bulk content generation, translation or data cleanup Requests that don't finish in the window expire, and only completed requests get billed. Anthropic and Google sell the same 50%-off batch deal on their APIs, so this lever travels with you if you switch providers. 5. Use Flex processing for traffic in between Some work is too slow for real time but can't wait a day either. Flex processing sits there. It bills at the same 50% discount as Batch, but you call it synchronously: set service_tier: "flex" on the request. The trade-off is availability. When capacity is tight, Flex can return a 429 error instead of an answer (those failed calls aren't charged). Wrap it in a retry with backoff, and fall back to Standard if a job is getting close to its deadline. The opposite setting deserves a warning too. "Fast" mode (the priority tier) costs 2x Standard on most models, and Ultrafast costs 6x. Turn those on only for the endpoints where latency clearly earns money, never as a global default. 6. Cap output and reasoning tokens Since output costs 5x input, a chatty model is an expensive model. Three fixes: Set max_output_tokens on every call. It includes reasoning tokens, so it also stops a reasoning model from thinking for 20,000 tokens about a yes/no question. Lower reasoning effort for simple tasks. Most extraction and classification work runs fine on low effort. Ask for structured output. A JSON schema with the exact fields you need cuts the preamble, the restated question and the "I hope this helps" at the end. A plain instruction like "Answer in under 80 words" also works better than people expect. Test it against your eval set before trusting it. 7. Send less context Input tokens are cheaper than output, but you send a lot more of them. Common leaks: Full chat history on every turn. After 20 turns you're paying for the whole transcript each time. Too many RAG chunks. Retrieving 20 chunks when the answer lives in 3 multiplies input cost with no quality gain. Tune top-k and add a reranker. (Our RAG chatbot tutorial covers this setup.) Whole documents where a section would do. There's a tension here worth knowing about. We used to summarize old conversation turns aggressively to save input tokens. Then we noticed summarization rewrites the start of the prompt, which kills the cache. Now the default is: keep the prefix stable as long as the conversation is short, and compact only when the history gets long enough that the savings clearly beat the lost cache hits. One more threshold: on current flagship models, requests over 272K input tokens switch to long-context pricing, which doubles the input rate for the entire request. If you're near that line, trim to stay under it. 8. Watch the hidden line items Token counts aren't the whole bill. Check your usage dashboard for these: Web search tool: $10 per 1,000 calls, plus the search content tokens at model rates. File search tool: $2.50 per 1,000 calls, plus storage at $0.10 per GB per day after the first free GB. Containers (Code Interpreter and hosted shell): billed per 20-minute session, from $0.03 to $1.92 depending on memory size. Regional data residency: a 10% uplift on newer models. An agent that runs a web search on every step can quietly spend more on tool calls than on tokens. 9. Put hard limits on agents Agents are where bills explode, because one bad loop can make hundreds of calls before anyone notices. Build the brakes into your own code: A maximum number of steps per task A per-task and per-user token budget, checked before each call A stop rule when the agent repeats the same error or the same tool call twice Separate API projects per environment, so a runaway test in staging can't touch production spend Project budgets in the OpenAI dashboard are useful for alerts, but don't count on them to stop traffic. The hard stop belongs in your application. If you're scoping an agent project, our breakdown of AI agent pricing in 2026 covers how running costs fit into the total. How much can you cut your OpenAI API bill? A worked example Here's the math for a hypothetical support assistant, using OpenAI's list prices. It handles 100,000 requests a month. Each request sends 3,000 input tokens (2,500 of them a fixed system prompt and knowledge snippet) and gets back 400 output tokens. Step Monthly cost Everything on GPT-6 Astra $5,000 Route to GPT-6.1 Sol after it passes evals $1,000 Fixed 2,500-token prefix served from cache ~$531 Output capped at 250 tokens ~$381 That's about a 92% cut, before touching Batch or Flex at all. Your numbers will differ. The order of the savings usually won't: model choice first, caching second, output discipline third. Where to start this week Lever Typical saving Effort Cost logging per feature Shows where to act Low Smaller model for simple tasks Often 80%+ on that feature Medium (needs evals) Cache-friendly prompt order Up to 90% on repeated input Low Batch API for async jobs 50% Low to medium Flex for in-between traffic 50% Low (add retries) Output and reasoning caps Varies, often large Low Agent step and budget limits Prevents blowups Medium If you only do two things to cut your OpenAI API bill, add per-feature cost logging today and reorder your prompts for caching tomorrow. Both take hours, not weeks, and they tell you exactly where the remaining money is going. And if you'd rather have someone do the audit for you, Craxinno builds and tunes production AI features on the OpenAI and Claude APIs. Tell us what you're running and we'll show you where the bill can come down.



