The model behind an interface
One module owns every model call, with provider and model in configuration. Swapping vendors is a setting, not a rewrite.
App costs · 2–8 weeks
How much does an AI app cost to build in 2026?
An app with a model in it costs what the app costs, plus the work of making the model reliable. A single AI feature inside a web app is a $5K MVP Sprint. A product whose core job is the model, across web and mobile, is a Product Build, $10K–$20K. Here is where the line is.
Typical price
$5K–$20K
Time
2–8 weeks
Platforms
Web, or web + iOS + Android
Where the number comes from
Calling a model is cheap and quick. Shipping one is not, because the model is wrong some of the time and the product has to be right. The cost lives in the parts around the call: the prompt and the examples, the structure forced on the output, the retries and fallbacks, the limits that stop one user burning the month's budget, and the evaluation set that tells you whether a change made things better. Resumemaster rewrites résumés for applicant tracking systems; the rewriting is a few lines, the checking is most of the code.
In the MVP Sprint the model does one bounded job inside an otherwise ordinary web app: summarise, classify, extract, draft, answer over a fixed set of documents. The input and output are defined and the failure mode is a polite retry. In the Product Build band the model is the product: a coach that edits a plan, as in Month6, retrieval over a customer's own data, multi-step tasks, voice, or a mobile app whose every screen depends on it.
Model cost is a running expense, not a build cost, and it sits in your account from day one. The build sets it up so you can see cost per user, cap it, and swap the model when a better or cheaper one appears. No model name is baked in; the interface is.
What the price includes
The scope, as it would be written.
Structured, validated output
The model is asked for a schema and the answer is checked against it. Bad output is retried or rejected; the user never sees raw text that should have been a form.
Retrieval where it is needed
Your documents chunked, embedded and searched in PostgreSQL with pgvector, so answers cite your content and not the model's memory.
Cost and rate limits
Per-user and per-day caps, token accounting in the database, and a dashboard line for what each feature costs to run.
An evaluation set
Thirty to a hundred real examples with expected results, run before every change to the prompt or model. The only honest way to know it still works.
What moves the price
Four things that change the number.
Agents and tools
A model that takes actions, calls APIs or runs multi-step plans needs guardrails, approval steps and a lot of testing. A Product Build, always.
Voice and images
Speech in and out, or image understanding, each add a modality with its own latency and cost profile.
Fine-tuning
Rarely the right first move. Prompting plus retrieval covers most products, and the evaluation set tells you when it stops being enough.
Compliance
Health, finance or legal data constrains where the model runs and what is logged, and can rule out some providers. Settled before the price is fixed.
Built here
Resumemaster
AI résumé engine · Next.js + React Native · Web + mobile
The rewriting is a few lines; the checking of what comes back is most of the code.
Questions
What people ask about ai app costs.
- Can an AI feature really fit in $5K?
- Yes, if it is one bounded job in a web app: summarise this, extract that, answer from these documents. The MVP Sprint covers auth, the feature with validation and limits, Stripe and deployment. It does not cover agents, voice, mobile apps or retrieval over large private datasets.
- Which model will you use?
- The one that fits the job and the budget at the time, kept behind an interface so it can change. Model names are not written into the product or into this page, because they go out of date faster than the code.
- What will the model cost me each month?
- It depends on usage, and the build makes it visible: token accounting per feature and per user, with caps. For most bounded features it is a small fraction of hosting. Agents and long documents cost more, which is one reason they are scoped carefully.
- How do you make sure it is accurate?
- Structured output that is validated, retrieval so answers come from your data, and an evaluation set of real examples run on every change. Where accuracy matters, the design keeps a human in the loop, as Month6's coach does by refusing medical advice.
Already built it?
If the ai app exists already, built in Lovable, Bolt, Cursor or Replit, the number on this page is not yours: the $1.5K fixed Production Audit reads that ai app against the published checklist and quotes its fixes in five days.
Your number