---
title: "App Like ChatGPT: Models, Costs and the First Version"
author: "Nate Laquis"
author_role: "Founder & CEO"
date: "2026-11-18"
category: "How to Build"
tags:
  - app like chatgpt
  - build an ai assistant app
  - llm api pricing
  - ai chat app development
  - ai app unit economics
excerpt: "You do not train a model to build an app like ChatGPT. You build a product on a model API, and a focused first version runs roughly $60K to $130K over 3 to 4 months. Here are the parts, the per-user math at current API prices, and the order to build in."
reading_time: "13 min read"
canonical_url: "https://kanopylabs.com/blog/app-like-chatgpt-models-costs-and-the-first-version"
---

# App Like ChatGPT: Models, Costs and the First Version

## What an App Like ChatGPT Takes

To build an app like ChatGPT you do not train a model. You build a product on a model API from OpenAI, Anthropic or Google and pay per token. A focused first version has streaming chat, saved conversations, sign-in, answers grounded in your own data, usage limits and a subscription, and on the open market it runs roughly $60,000 to $130,000 over 3 to 4 months.

The model is a line item you can swap in an afternoon. The product is everything around it: what the assistant knows that the general one does not, what it can do besides talk, and what each user costs you a month.

This article covers those decisions for a founder planning an assistant for one group of people: lawyers, students, electricians, a company's own staff, or someone who wants a companion. If your plan is a support bot on your website, that is a different build, and our guide on [how to build an AI chatbot](/blog/how-to-build-an-ai-chatbot) covers it.

![Close-up of two hands typing a message on a smartphone keyboard](https://images.unsplash.com/photo-1526045612212-70caf35c14df?w=800&q=80)

## Rent the Model, Pick a Narrow Job

Training a frontier model costs more than your company will raise, so you rent one. There are four places to rent from. OpenAI, Anthropic and Google each sell a flagship, a mid-tier model and a cheap fast one. The fourth is an open-weight model run by a host such as Together AI, Fireworks or Amazon Bedrock, which you pick when you need the weights under your control, a fine-tune of your own, or a contract a hospital's lawyers will sign. Start with one of the big three. Their mid-tier models are good enough for almost every assistant product.

The harder decision is the job. A general assistant that answers anything competes with ChatGPT, Claude and Gemini, which are already on your customer's phone and improve every few months without your help.

A narrow assistant wins on things the general one does not have:

- **Your data.** State case law with citations a lawyer can check. The actual curriculum a school district teaches. The electrical code edition a county enforces, plus the manufacturer manuals.

- **Actions.** It drafts the filing in the firm's template, books the tutoring session, or builds the parts list and sends it to the supplier.

- **A workflow.** The first screen asks about the case, the lesson or the job site. A blank text box asks the user to do the product design for you.

- **Trust in one field.** Sources on every answer, a refusal when the documents do not cover the question, and review by a person where the stakes call for it.

We saw the pattern on [Politik](/work/politik), a politics app we built in React Native with three first-time founders. Politik AI is a chat assistant grounded in official data that explains any bill, vote or process in plain English. The app took under three months to reach the App Store and had 4,000 users on launch day.

## The Parts of the Product

The model call is about twenty lines of code. The time goes into the pieces around it.

### Chat, history and accounts

Answers stream in word by word, because a four-second wait on a blank screen feels broken and the same four seconds of text appearing feels fast. Streaming brings its own work: a stop button, retry, a dropped connection mid-answer, and markdown that renders while it is still arriving. Conversations are saved per user with generated titles, search and delete. Sign-in is email code, Apple and Google through a service such as Clerk or Supabase Auth.

### Model routing

Put a thin layer between your app and the vendors. It decides which model answers each message: the cheap one for small talk, titles and short follow-ups, the mid-tier one for the work your users pay for. It also retries on a second vendor when the first has an outage. Nothing else you build moves your margin as much.

### Retrieval over your own data

Retrieval-augmented generation, or RAG, is how the assistant answers from your documents. You split them into chunks, store an embedding for each chunk, find the closest ones to the question, and hand them to the model with an instruction to cite. Embedding is cheap. OpenAI lists text-embedding-3-small at $0.02 per million tokens, so a 50-million-token library costs about a dollar to index. The cost is in the pipeline: parsing ugly PDFs, keeping documents current, and testing whether the right chunk comes back.

### Tools and memory

Function calling lets the model ask your backend to do something: look up an order, search the web, run a calculation, create a calendar event. Each tool is a small API with its own permission check, and anything that spends money or sends a message should ask the user to confirm first. Memory is a short profile of facts the assistant keeps between conversations, such as a student's grade level or a contractor's license state. Show users what is stored and let them delete it.

### Files, images and voice

Current models read photos and PDFs directly, so a user can photograph a breaker panel or upload a contract. You still build the upload, the storage and the size limits. Voice comes in two grades. Dictation, where speech becomes text and the answer is read aloud, is cheap: OpenAI lists gpt-transcribe at about half a cent per minute. A live voice conversation is a different bill. OpenAI lists GPT-Live-1 voice sessions at $0.05 a minute, which puts a user who talks ten minutes a day at about $15 a month before any text.

### Analytics and evals

Log every request with the user, the model, the tokens and the latency. Add thumbs up and down on answers. Then keep an eval set: 100 to 200 real questions with answers a domain expert approved, run against every prompt change and every new model. Without it, a one-star review is how you learn an update made answers worse.

## What One User Costs You a Month

Do this math before you pick a price. The figures below are list prices per million tokens from the vendors' pricing pages, checked on November 18, 2026. They change several times a year.

- **Flagship:** OpenAI's gpt-6.2-astra and Anthropic's Claude Fable 5.2 both list at $8 input and $40 output. Google's Gemini 4 Carbon lists at $2 and $10 through November 30, then $4 and $20 from December 1.

- **Mid-tier:** gpt-6.2-sol and Claude Sonnet 5.5 both list at $2 input and $10 output. Gemini 4 Hydrogen is $1.50 and $7.50.

- **Cheap and fast:** gpt-6-luna is $0.10 and $0.50. Claude Haiku 5.5 is the same for prompts up to 100,000 tokens. Gemini 2.5 Flash-Lite is $0.10 and $0.40.

A token is about three quarters of an English word. The part founders miss is that every message resends everything: your system prompt, the conversation so far, the retrieved documents and the new question. For an assistant with retrieval, plan on an average of 6,000 input tokens and 500 output tokens per message.

On a $2 and $10 model that is 1.2 cents of input and half a cent of output, so 1.7 cents a message. On a $0.10 and $0.50 model it is under a tenth of a cent. On an $8 and $40 flagship it is 6.8 cents.

### Three users

- **Light, 60 messages a month:** about $1 on a mid-tier model and 5 cents on a cheap one.

- **Typical, 300 messages a month:** about $5.10 on mid-tier, 26 cents on cheap, about $20 on a flagship.

- **Heavy, 1,500 messages a month:** $25.50 on mid-tier and $102 on a flagship. That is 50 messages a day, and every assistant has a few.

Now the price. A $20 subscription sold through the App Store nets you $14 to $17 after Apple's 15 to 30 percent. The typical user leaves a healthy margin on a mid-tier model. The heavy user costs more than they pay, and each one cancels the profit from a typical subscriber.

Routing and caching fix most of it. Send 70 percent of messages to the cheap model and the typical user drops to about $1.70 a month, the heavy one to about $8.50. Prompt caching cuts the repeated part of each request further: both $2 models bill cached input at $0.10 per million, a twentieth of the normal rate. A fair target is model spend under a quarter of what you net per subscriber.

Free users are the other half of the bill. Ten thousand free users sending 60 messages a month cost about $510 on a cheap model and about $10,200 on a mid-tier one. Free tiers run on the cheap model, with a daily cap. For help choosing among the three vendors, read our comparison of [Claude, GPT and Gemini for apps](/blog/claude-vs-gpt-vs-gemini-for-apps).

![Person in a checked shirt pressing keys on a blue desk calculator beside a keyboard](https://images.unsplash.com/photo-1626266061368-46a8f578ddd6?w=800&q=80)

## Limits, Metering and Getting Paid

Flat pricing with unlimited use is how assistant apps lose money. Meter from the first day, even if you never show the user a number.

### Metering and rate limits

Write a usage row for every request: user, model, input tokens, output tokens, cost. Limits read from that table. The pattern that works is a daily cap on the good model with a quiet fallback to the cheap one, so a heavy user's answers get a little plainer instead of stopping. Add a per-minute limit per account and per device to stop scripts, and a hard monthly spend alarm on your vendor account. Someone will find your endpoint and try to use it as a free API.

Vendors rate-limit you too. New accounts start on low tiers that rise with spend, so ask for an increase a few weeks before launch.

### App Store billing or Stripe

On iOS, Apple's guideline 3.1.1 says paid features inside the app go through in-app purchase, and Apple keeps 15 to 30 percent. Google Play has the same kind of rule and a similar cut. On the web you take cards through Stripe at roughly 2.9 percent plus 30 cents in the US. On a $20 plan that is about $0.88 to Stripe against $3 to $6 to Apple.

So sell on both. Put a web checkout on your site and send your own traffic there. Offer in-app purchase in the mobile apps, because people who find you in the store will pay there or not at all. A service such as RevenueCat keeps one subscription record across Apple, Google and Stripe so the backend asks a single question: is this user paid.

In the United States storefront, Apple's guidelines now let an app link out to a web checkout without a special entitlement. What Apple may charge on those purchases is still being argued in court, so check where it stands before your pricing depends on it.

### Plans

Most consumer assistants land between $10 and $25 a month, with an annual plan at about two months free. Professional tools for lawyers, clinicians or engineers charge $50 to $200 a seat, and at that price the token bill stops mattering. Credits fit products where one request is expensive, such as long document review or live voice.

## Safety, App Review and Privacy

An app that generates text for strangers gets a closer look in review than a to-do list does. Build these in from the start.

### Moderation

Check what goes in and what comes out. The vendors offer moderation models, and you add your own rules for your field: a tutoring app for kids blocks different things than a legal research tool. Test the system prompt against the jailbreak prompts that circulate online, and answer self-harm messages with a fixed response that points to real help. Companion apps need this more than any other category.

### What app review checks

- **Guideline 1.2.** Apple groups AI content creation apps with user-generated content. It asks for a filter for objectionable material, a way to report it, the ability to block abusive users, and published contact details. Put a report button on every answer.

- **Guideline 5.1.2(i).** You must disclose where personal data is shared with third parties, including third-party AI, and get explicit permission first. That means a consent screen naming your model vendor before the first message.

- **Guideline 4.3(b).** Apple rejects apps that are indistinguishable from what is already widely available. A model with a new logo on it is the textbook case.

- **Account deletion.** If users can create an account, they must be able to delete it inside the app.

- **Age rating.** Answer the questionnaire for the worst thing the model could say. A mis-rated chat app can be pulled.

Google Play's AI-generated content policy covers chatbot apps where the chat is the central feature, and it holds you responsible for what the model generates.

### Privacy

People tell an assistant things they would never type into a search box. Read each vendor's data terms before you sign up: whether API traffic is used for training, how long prompts are kept, and whether zero retention is available to you. Then encrypt conversations at rest, set a retention period, let users export and delete their history, and keep raw prompts out of your analytics tools.

Regulated fields add paperwork. Health data in the US needs a business associate agreement with every vendor that touches it. A product for children under 13 falls under COPPA and needs verifiable parental consent. For a legal assistant, nobody on your team reads a conversation unless the user asks.

## The Stack We Would Pick

Nearly all of this is ordinary product engineering. Save the experiments for retrieval and prompts.

- **Mobile:** React Native with Expo, one TypeScript codebase for iOS and Android.

- **Web:** Next.js for the web app, the marketing site and the checkout.

- **Model layer:** The Vercel AI SDK if you want one interface over several vendors with streaming and tool calls handled. The vendor SDKs directly if you are committed to one provider and want its newest features on release day.

- **Database:** Postgres for users, conversations, messages and usage, with the pgvector extension for embeddings. One database covers you into the millions of chunks.

- **Background work:** A job queue for document ingestion, title generation, memory updates and eval runs.

- **Vendors:** Clerk or Supabase Auth, Stripe and RevenueCat for billing, object storage for uploads, and a tracing tool such as Langfuse or LangSmith for prompts and costs.

Two rules matter more than the list. Never call the model from the phone, because anyone who wants the API key inside a shipped app can pull it out. Every request goes through your backend, which checks the user, applies the limits and records the usage. And keep prompts and model names in configuration, so changing either is a deploy of one file and an eval run.

![Laptop screen in a dark room showing colorful lines of code in an editor](https://images.unsplash.com/photo-1607799279861-4dd421887fb3?w=800&q=80)

## Cost and Timeline

These are market ranges for a senior product team in the US or Western Europe. Scope moves them more than hourly rates do, and our [app development cost](/app-development-cost) page covers what drives app budgets in general.

- **Focused first version:** $60,000 to $130,000, 3 to 4 months. A web app or one React Native app, streaming chat, history, sign-in, retrieval over one body of documents, two or three tools, usage limits, and one paid plan.

- **Full product on web and mobile:** $130,000 to $280,000, 5 to 8 months. Adds the second platform, file and image input, memory, dictation, model routing with a fallback vendor, an admin console, a real eval suite, and billing on the App Store, Google Play and the web.

- **Platform:** $280,000 to $500,000 and up, 9 to 12 months. Live voice, team workspaces with roles, enterprise sign-on, audit logs, compliance work for a regulated field, and a fine-tuned or self-hosted model where the contracts demand it.

Four things move the number. The first is the state of your data: clean, licensed documents are weeks of work and a pile of scanned PDFs is months. After that it is how many tools the assistant gets, whether you need live voice, and compliance.

After launch you have two running costs. Maintenance is typically 15 to 20 percent of the build cost per year. The model bill scales with use: at 1,000 paying users on the routed setup above, expect roughly $1,500 to $5,000 a month in tokens, plus $300 to $1,500 for hosting and the other vendors. For a line-by-line view of chatbot budgets, see [how much it costs to build an AI chatbot](/blog/how-much-does-it-cost-to-build-an-ai-chatbot).

## What to Build First

Before any app, spend two weeks on the answer. Take 50 real questions from the people you are building for, your documents, and a mid-tier model in a script or a notebook. If a domain expert reads the output and would not pay for it, no interface will save the product. If they would pay, those 50 questions start your eval set.

### The order to build in

1. **Retrieval and the eval set.** Ingestion for your first body of documents, citations on every answer, and 100 questions with approved answers.

2. **The backend chat endpoint** with streaming, usage logging and per-user limits.

3. **The chat screen and history** on one platform. Pick web if your users sit at desks, mobile if they are on a job site or a couch.

4. **Accounts, the consent screen and the paid plan.** Charge from the first release. Free pilots tell you people like free things.

5. **One or two tools** that finish the job the answer starts: export the draft, book the session, send the parts list.

6. **Routing to a cheap model,** once the logs show which messages do not need the good one.

7. Later: the second platform, memory, file and image input, voice, and team accounts.

That order gets a paid product in front of real users at about week ten. Their conversations decide what comes next.

We build apps like this. Bring the audience, the documents or data your assistant should know, and the one job it has to do well. We will scope the first version, pick the models, run the per-user math with you, and give you a plan and a number. [Book a free strategy call](/get-started)

---

*Originally published on [Kanopy Labs](https://kanopylabs.com/blog/app-like-chatgpt-models-costs-and-the-first-version)*
