Florin Florea··11 min read

AI App Development Cost Calculator: Build & API Bands

AI app development cost by scope: web app build bands from the engine, model API prices from vendor pages, data and evaluation work, and running cost.

FF

Florin Florea

10+ years web dev · Scoping web projects since 2011

Some links in this guide are affiliate links — if you buy through one, I earn a commission at no extra cost to you. I only link to tools I'd use myself.

Want your specific number? Try our free calculator — it takes 2 minutes.

Open the Free Cost Calculator

Quick answer: what an AI app costs to build

An AI app costs $16,100–$32,200 with a freelancer or $27,600–$55,250 with an agency when the AI is one feature on a simple web app, $77,300–$154,550 or $132,700–$265,450 as a full product with dashboards and integrations, and $154,550–$309,050 or $265,500–$530,950 when it needs real-time behaviour and its own AI pipeline. Those are the engine's web app bands for the US market; the medians for the top scope are $220,750 (freelancer) and $379,250 (agency). Source: Project Cost Estimator engine, calibrated against 600+ project quotes and public rate benchmarks across 6 markets.

The build is only part of the bill. An AI app also pays per use for the model behind it, and that line is priced by the model vendor, not by the developer. This page keeps the two apart: build cost from the engine, model cost from the vendors' own pricing pages, quoted as published and linked in the sources. Anything this page cannot source to one of those two places is described without a number.

The calculator above runs the web app niche, which is where the engine prices AI work. For a native mobile app with an AI feature, use the US app calculator and read this page for the model line.

Web App Website Cost Calculator

Live estimate

What are you building?

Where's your developer?

Budget

$16,100

Realistic

$23,000

Premium

$32,200

~5.5 weeksCustom Development (React/Next.js/Node) recommended

Email me this estimate — your reference number when the quotes come in

No spam, ever. Unsubscribe anytime.

Powered by the same 9-engine calculator behind every estimate on this site — full version →

What counts as an AI app, and which scope it lands in

Most AI apps built for a business are one of three shapes. Each maps to one of the engine's web app scopes.

  1. 1. A model feature inside an app. Summarise this document, draft this reply, classify this ticket. The app calls a hosted model through an API, shows the result, and stores it. The AI part is a few hundred lines; the app around it is a normal web app. Scope: MVP / simple CRUD.
  2. A product built on your own data. A support assistant that answers from your help centre, a search tool over contracts, an internal analyst over your reports. This needs an ingestion pipeline, an embedding and retrieval layer, a way to show sources, and dashboards for whoever owns it. Scope: full product (dashboards, integrations).
  3. An agent that acts. It takes a goal, calls tools (your CRM, your calendar, a payment system), checks results and tries again. It needs real-time behaviour, permissions, an audit trail, and evaluation before anyone trusts it. Scope: enterprise (real-time, AI).

Outside the bands: training a model from scratch, or fine-tuning at a scale that needs its own GPU budget. Those are research projects with hardware lines, and no calculator built on project quotes prices them. Everything below assumes a hosted model from a vendor.

AI app build cost by scope: engine web app bands

Engine output for the web app niche, US market. Weeks are elapsed calendar weeks for one focused build.

ScopeFreelancer band (median)WeeksAgency band (median)Weeks
MVP / simple CRUD (one model feature)$16,100–$32,200 ($23,000)~5.5$27,600–$55,250 ($39,450)~4
Full product (dashboards, integrations; retrieval over your data)$77,300–$154,550 ($110,400)~24.5$132,700–$265,450 ($189,600)~16
Enterprise (real-time, AI; agents with tools)$154,550–$309,050 ($220,750)~50.5$265,500–$530,950 ($379,250)~33.5


Why the AI scopes are not more expensive than the equivalent non-AI scopes: the hard part of a hosted-model app is the app. Calling a model is one HTTP request. Getting good results out of it is prompt design, retrieval and evaluation, and those are hours of ordinary engineering that the engine already counts.

Why they are not cheaper either: an AI feature fails in ways a form does not. It answers wrongly with confidence. The work that stops that from reaching a user (evaluation sets, guardrails, a way to show sources, a fallback when the model is down) is where the hours go, and a quote that skips it is cheaper because it is unfinished.

The enterprise AI scope in six markets

The top scope, real-time and AI, by where the team sits. Same build, different rates.

MarketFreelancer band (median)Agency band (median)
United States$154,550–$309,050 ($220,750)$265,500–$530,950 ($379,250)
United Kingdom$137,400–$274,750 ($196,250)$231,150–$462,350 ($330,250)
Australia$137,400–$274,750 ($196,250)$231,150–$462,350 ($330,250)
Western Europe$120,200–$240,450 ($171,750)$211,700–$423,450 ($302,450)
Eastern Europe$59,600–$119,200 ($85,150)$120,200–$240,450 ($171,750)
South Asia$35,550–$71,100 ($50,800)$72,150–$144,350 ($103,100)


One caution that applies more to AI work than to a standard web app: the model vendor's bill does not move with the team's location. A cheap build with an expensive model in the loop can cost more over a year than the reverse. Price the two lines separately and add them.

The AI-specific lines: what is priced and what is not

Model API usage. Hosted models are billed per token, with separate prices for input (what you send) and output (what comes back), per million tokens. From the vendors' published pricing pages on the day this page was written, standard tier: OpenAI lists gpt-5.4-mini at $0.75 input and $4.50 output per 1M tokens, and gpt-5.5 at $5.00 input and $30.00 output; its text-embedding-3-small is $0.02 per 1M input tokens. Google lists Gemini 3.1 Flash-Lite at $0.25 input and $1.50 output per 1M tokens, and Gemini Embedding 2 at $0.20 per 1M text tokens. Vendor pages change; the links in the sources are the live numbers.

A usage example, arithmetic only. An app that sends 1,000,000 input tokens and gets back 200,000 output tokens a day on gpt-5.4-mini pays $0.75 plus $0.90, which is $1.65 a day, or $49.50 over thirty days. The same traffic on gpt-5.5 is $5.00 plus $6.00, which is $11 a day, or $330 a month. On Gemini 3.1 Flash-Lite it is $0.25 plus $0.30, which is $0.55 a day, or $16.50 a month. The model choice moves the running cost by a factor of twenty on the same traffic, which is why it is a product decision, not a developer default.

Embeddings and retrieval. A product over your own data embeds every document once and every query as it arrives. The embedding prices above are per million tokens; a vector store to hold them is a separate service with its own pricing, which this page does not quote because no vendor page for it was fetched. Ask the developer which store they propose and read that vendor's page.

Data work. Cleaning and structuring the documents the model will read, deduplicating, deciding what is out of scope. It is billed at the developer's hourly rate and it is the line most often missing from a first quote. There is no vendor price for it; it is hours.

Evaluation. A set of real questions with known good answers, run against every change. Without it nobody can say whether a prompt change made the product better or worse. Hours again, and they recur.

Fine-tuning. Vendors offer it at per-token training prices on the same pricing pages linked below. It is rarely the right first step; most products get further with retrieval and prompt work before they touch it.

Guardrails and fallbacks. Input filtering, output checks, rate limits, a graceful answer when the model is unavailable. Ordinary engineering, inside the build band.

Monthly running cost: hosting plus model usage

The engine prices running costs for the web app niche at $176 a month for the MVP scope, $226 for the full product scope and $386 for the enterprise scope. That is hosting, backend services, monitoring and tooling.

Model usage sits on top and scales with users, not with code. The example above puts the same daily traffic between $16.50 and $330 a month depending on the model. Multiply by your expected traffic, and then check it against the vendor's page again before launch, because the prices move and the newest model is rarely the cheapest.

Two habits keep the model line under control. Route simple requests to a small model and only hard ones to a large one. And cache: the same question asked twice should hit the model once. Both are a few days of engineering and both pay back in the first month at any real volume.

Freelancer vs agency for AI work

Hourly figures are the engine's rate table, custom development row, US column: $88–$218 an hour for a freelancer (median $135) and $145–$385 for an agency (median $232).

A freelancer with real experience of hosted-model products is the fastest route to the first scope: one feature, one model, a few weeks. The risk is evaluation. One person tends to test on the examples they wrote, and ships a feature that works on those.

An agency is the better fit for the second and third scopes, because retrieval products and agents need a designer for the trust surfaces (sources, confidence, what the system did), a tester who builds the evaluation set, and someone who owns the model bill. Ask any agency quoting AI work three questions: which model and why, how they will measure whether answers are right, and what the fallback is when the model is down or wrong. The quote that answers all three is the one to compare with the band.

Both will quote the build. Neither will quote the model usage, because it depends on your traffic. Ask for the per-request token estimate and do the arithmetic yourself from the vendor page.

How long an AI app takes to build

The engine's weeks: about 5.5 (freelancer) or 4 (agency) for the one-feature scope, about 24.5 or 16 for a full product, and about 50.5 or 33.5 for an enterprise build.

The AI-specific additions to the calendar:

  1. 1. Data access. Getting the documents, the exports or the API keys from inside your organisation. This is the most common cause of a late start, and it is on your side.
  2. Evaluation rounds. Each round of prompt or retrieval changes needs a run against the evaluation set and a human look at the failures. Budget rounds, not days.
  3. Vendor review and limits. New API accounts carry usage limits that lift with spend history. A launch that expects heavy traffic on day one needs the limits raised in advance.
  4. The trust surface. Showing sources, letting users correct answers, logging what the system did. It is not optional for anything that acts on a user's behalf, and it is design work as much as code.

How to use the calculator

1. Pick the scope by shape: one model feature is the MVP band; a product over your own data is the full product band; an agent that acts is the enterprise band.

  1. 2. Choose freelancer or agency. For the second and third scopes the agency column is usually the honest one.
  2. Switch the market to where the team sits.
  3. Add the model line separately: estimate tokens per request, multiply by requests per day, price it from the vendor page, and put the monthly figure next to the engine's running cost.

Related calculators: the web app development cost page runs the same engine with the full feature list; the SaaS cost estimator prices a subscription product end to end; the MVP cost calculator prices a first version by scope. Hourly rates by region are in the rate database, and how the bands are built is on the methodology page. For the hourly arithmetic, see the app development cost per hour calculator.

Sources

1. Project Cost Estimator engine, calibrated against 600+ project quotes and public rate benchmarks across 6 markets. Method: https://projectcostestimator.com/methodology

  1. 2. Hourly rate table by region and platform; the custom development row is used for app work: https://projectcostestimator.com/rate-database (source: Clutch Apr 2026, Talmatic Global Snapshot, Arc.dev, Upwork medians, TechReviewer, calibrated to real 2026 agency and freelancer benchmarks)
  2. OpenAI API pricing, standard tier, per 1M tokens, as published on the day this page was written: https://developers.openai.com/api/docs/pricing
  3. Google Gemini API pricing, paid tier, per 1M tokens, as published on the day this page was written: https://aistudio.google.com/docs/pricing
  4. Web app calculator: https://projectcostestimator.com/web-app-development-cost

Engine figures used (web app niche, US dollars). US market: MVP / simple CRUD freelancer $16,100–$32,200 (median $23,000), agency $27,600–$55,250 (median $39,450); full product freelancer $77,300–$154,550 (median $110,400), agency $132,700–$265,450 (median $189,600); enterprise (real-time, AI) freelancer $154,550–$309,050 (median $220,750), agency $265,500–$530,950 (median $379,250); monthly running cost $176, $226 and $386. Enterprise scope in other markets: UK and Australia freelancer $137,400–$274,750 (median $196,250), agency $231,150–$462,350 (median $330,250); Western Europe freelancer $120,200–$240,450 (median $171,750), agency $211,700–$423,450 (median $302,450); Eastern Europe freelancer $59,600–$119,200 (median $85,150), agency $120,200–$240,450 (median $171,750); South Asia freelancer $35,550–$71,100 (median $50,800), agency $72,150–$144,350 (median $103,100). US hourly rate for custom development: freelancer $88–$218 (median $135), agency $145–$385 (median $232).

Vendor figures quoted (per 1M tokens, standard or paid tier, from the pages above): OpenAI gpt-5.4-mini $0.75 input, $4.50 output; gpt-5.5 $5.00 input, $30.00 output; text-embedding-3-small $0.02 input. Google Gemini 3.1 Flash-Lite $0.25 input, $1.50 output; Gemini Embedding 2 $0.20 text input. Usage arithmetic for 1,000,000 input and 200,000 output tokens a day: gpt-5.4-mini $0.75 plus $0.90 is $1.65 a day, $49.50 over thirty days; gpt-5.5 $5.00 plus $6.00 is $11 a day, $330 a month; Gemini 3.1 Flash-Lite $0.25 plus $0.30 is $0.55 a day, $16.50 a month.

Ready to build? Three ways to do it

The right path depends on budget, not preference. These bands come from our own published cost data.

Do it yourself

$0–$500

WordPress on cheap hosting like SiteGround — enough for a simple site you maintain yourself.

Hire a freelancer

$500–$5,000

Custom design and setup without agency overhead — the sweet spot for most small-business sites.

Browse Fiverrdevelopers →

Custom project

$5,000+

Scope it properly before you talk to anyone — our calculator prices your exact feature list so quotes can't inflate.

Price my project →

Magento & Shopify Plus

Scopebit

The agency behind this calculator builds, upgrades and migrates Magento and Shopify Plus stores for companies across Europe and North America.

Book a call with Scopebit →

Disclosure: some links above are partner links — we may earn a commission at no extra cost to you. Recommendations come from our cost data, not from who pays us.

Get your personalized estimate

Our 9-engine calculator analyzes 30+ features, platform-specific rates, and your geographic market.

Start Free Estimate

Free · No signup · Results in 2 minutes

Frequently Asked Questions

How much does it cost to build an AI app?+
From the engine web app bands, US market: $16,100–$32,200 (freelancer) or $27,600–$55,250 (agency) for one model feature on a simple app; $77,300–$154,550 or $132,700–$265,450 for a product over your own data with dashboards; $154,550–$309,050 or $265,500–$530,950 for an agent with real-time behaviour. Model usage is billed separately by the vendor.
What does the model API cost per month?+
It depends on traffic and model. From the vendor pages quoted on this page, 1,000,000 input and 200,000 output tokens a day costs $49.50 a month on gpt-5.4-mini, $330 on gpt-5.5 and $16.50 on Gemini 3.1 Flash-Lite. Estimate your tokens per request, multiply, and check the vendor page before launch.
Is an AI app more expensive than a normal web app?+
Not at the build stage. The engine prices AI work inside its web app scopes because calling a hosted model is ordinary engineering. The extra cost is the model usage line, the data preparation and the evaluation work, and those are the lines a cheap quote leaves out.
Does the calculator include fine-tuning or training a model?+
No. Fine-tuning is priced per token on the vendor pages linked in the sources and is rarely the right first step. Training a model from scratch is a research project with a hardware budget and no calculator built on project quotes prices it.
Should I use a freelancer or an agency for an AI product?+
A freelancer for a single model feature. An agency for a retrieval product or an agent, because those need design for the trust surfaces, a tester who builds the evaluation set, and someone who owns the model bill. Ask either which model, how they measure correctness, and what the fallback is.

Related Articles

ai app development costai development costai app development cost calculatorhow much does it cost to build an ai appcost to develop an ai app