How to Price an AI Wrapper Without Losing Money

The short answer

Price an AI wrapper on the job it does for a specific customer, never on the tokens it resells. Cost your heaviest workflow at the model's list price, sell plans with an included allowance and marked-up overage, and target at least 60% gross margin after inference. Then build a cost lever (routing, caching or your own model) and a moat the API vendor can't copy.

Key takeaways

A wrapper is only as profitable as the gap between your price and your model bill, and the vendor controls half of that gap.
Never sell unlimited access to a frontier model at a flat price. A single heavy user can cost 20 times their plan.
Price on the customer's job (a quote, a contract, a listing), not on tokens, so your price holds when model prices fall.
Overage billed at the API list price earns you nothing. Mark it up or route it to a cheaper model.
The fastest-growing wrappers still faced negative margins; the escape was owning a model or being owned by someone who does.

In the book, I warned about the dog biscuit problem.

Say you start selling dog biscuits on subscription. Great idea. But nothing stops me from doing the same thing. I could find a dog biscuit supplier, set up a template website overnight, and we're in competition.

Now swap the dog biscuits for an AI model.

Anyone can get an API key from OpenAI, Anthropic or Google, put a friendly screen on top, and launch an "AI tool for X" by Friday. That's an AI wrapper. And it has two problems the dog biscuit business never had.

First, your supplier sets your cost of goods on every single click. Second, your supplier sells the same thing to your customers directly, often for $20 a month.

So, can you build a profitable business on someone else's model? And more importantly, how do you price it so you don't lose money on your best customers? Let me show you.

What is an AI wrapper?

An AI wrapper is a product built on top of another company's AI model, usually reached through an API. You don't train the model. You add the interface, the instructions, the workflow and (if you're smart) the data and context the general model doesn't have.

"Wrapper" is often used as an insult. It shouldn't be. Most software is a wrapper around something. What matters is how thick the wrapper is.

Type	What you add	Example	How defensible
Thin prompt wrapper	A prompt and a nicer screen	A generic "chat with your PDF" tool	Weak: easy to copy, and models do it natively
Workflow wrapper	Steps, integrations and output in the format a job needs	Lovable turns a prompt into a working, hosted web app	Medium to strong
Data or context wrapper	Proprietary data, brand context, domain rules	Jasper's brand voices and knowledge assets	Medium
Vertical wrapper	Deep fit to one profession's work and risk	Harvey and Legora for lawyers	Strong, if priced against labor

Are AI wrappers profitable?

Honest answer? Some are growing faster than any software in history. Few are comfortably profitable yet.

Look at the evidence:

Cursor passed a reported $4 billion in annualized revenue by June 2026, according to Dealroom. Yet Contrary Research reported its gross margin at -23% for the quarter to January 2026. An earlier estimate put its Anthropic bill at about $650 million a year against about $500 million in revenue. According to Wikipedia's SpaceXAI entry, SpaceX agreed in June 2026 to buy Cursor for $60 billion, closing on 14 August 2026.
Lovable said it was tracking toward about $600 million ARR by the end of August 2026, according to Tech Startups. Sacra estimated its 2025 gross margin at about 35%.
Base44 was bought by Wix in June 2025 for about $80 million plus earn-outs. Wix later said Base44's gross margin was "nearly zero" at the start of 2026, and targeted about 60% after moving to its own model, according to Ynet.
Chegg isn't a wrapper, but it shows the other side. When free AI did its core job, Chegg's Q2 2026 revenue fell 51% to $51.8 million.

The pattern is clear. Revenue is easy to get. Margin is hard to keep. And the companies that escaped the squeeze either built their own model or were bought by someone who had one.

For the full benchmark table, see AI gross margins and inference costs.

How do AI wrappers make money?

Six ways, and most successful wrappers combine two or three:

Subscription with an allowance. A monthly plan that includes a set amount of usage, then charges for more. This is the 2026 default, and the AI pricing guide shows how to set one up.
Credits. A balance customers spend on actions of different sizes. Lovable charges 0.50 credits for a small edit and 1.70 for a landing page. See credit-based pricing.
Cost-plus resale. Publishing token rates with a markup. Vercel's v0 lists its own rates from $0.20 to $10 per million input tokens.
Per job or outcome. A price per quote, per contract review, per resolved ticket. See outcome-based pricing.
A premium seat sold against labor. Harvey is reported at $1,200 to $2,000-plus per seat per month because it's compared with lawyer time, not with ChatGPT.
A platform fee on top of pass-through. Clay passes token-heavy models through at cost but charges Actions for its own work.

Worked example: the wrapper that loses money and the one that doesn't

Let's run the numbers on two wrappers built on the same model, Claude Sonnet 5.5 ($2 per million input tokens, $10 per million output, checked 27 October 2026). The businesses are illustrative; the model prices are real.

Wrapper A: 'Chat with your documents', $10 a month, unlimited

A typical question sends 50,000 tokens of document context and gets a 500-token answer.

Cost per question: 50,000 × $2 ÷ 1M + 500 × $10 ÷ 1M = $0.10 + $0.005 = $0.105
A light user asking 40 questions a month costs $4.20. Fine.
A heavy user asking 2,000 questions a month costs $210, on a $10 plan.

One heavy user wipes out the profit from about 34 light ones. And the feature is something the model vendors include in their own $20 plans.

Wrapper B: 'QuoteBot' for trades businesses, $79 a month

A tradie records a three-minute voice note on site and gets a formatted, itemized quote.

Transcription: 3 minutes at Deepgram's $0.0043 per pre-recorded minute = $0.013
Drafting: 8,000 tokens in (price list, job template, transcript), 1,500 out = $0.016 + $0.015 = $0.031
Total: about $0.044 per quote
A typical user sending 120 quotes a month costs $5.28. A heavy user sending 400 costs $17.60, still a 78% margin before other costs.

And the price isn't anchored on tokens. It's anchored on the evening a tradie no longer spends writing quotes at the kitchen table.

Same model. Same supplier. The difference is that Wrapper B sells a job to a specific customer, needs no caps its customers would ever notice, and still can't be wrecked by its heaviest user.

AI Gross Margin CalculatorModel token costs, heavy users and margin for an AI product.Open the free tool →

Seven rules for pricing an AI wrapper

Rule 1: Sell the job, not the model

Your value metric should be something your customer already counts: quotes, listings, contracts, minutes of video. In the book I said a good usage metric must grow with your customers and communicate the value of your service. "Tokens" does neither for most buyers.

Rule 2: Never sell unlimited frontier access at a flat price

Cursor learned this in public in 2025. If you want the comfort of "unlimited," make the expensive work metered and keep a cheaper or slower mode unlimited, the way Midjourney does with Relax mode.

Rule 3: Cost at list price, then at tomorrow's price

Google has scheduled Gemini 4 Argon to double from $2/$10 to $4/$20 per million tokens on 1 December 2026. If your product runs on it, your pricing must already assume the higher rate. The LLM API pricing comparison flags every scheduled change.

Rule 4: Mark up your overage

Cursor and Devin charge usage beyond their plans at "API list price." For a wrapper that pays list price, that's zero margin on your heaviest users. Charge overage at a markup, or route it to a cheaper model.

Rule 5: Build a cost lever

Routing, caching, batching, a fine-tuned small model, or your own model. Cursor built Composer (see the Cursor pricing teardown). Base44 moved to its own model (see the Base44 teardown). Without a lever, every price war with the model vendor is lost before it starts.

Rule 6: Charge for your platform as well as the pass-through

If customers could buy the same tokens themselves, your resale margin is temporary. Clay split its pricing so customers who bring their own API keys still pay for the platform work. Price the workflow you own.

Rule 7: Dig a moat the API can't fill

In the book, MOATS stands for Marketing supremacy, One (no single point of failure), Audience, Toughness and Switching. A wrapper starts weak on two of them: it depends on one supplier, and its gross margin is thin. So build the other three. A loyal audience in one niche. Workflow and data that make switching painful. Distribution the vendor doesn't have, the way Base44 sits inside Wix. See moats in the AI era.

Is your wrapper making money on every customer?

Model your token cost per task, your heavy users and your plan price in the free AI Gross Margin Calculator, before a customer finds the gap for you.

Open the free tool

Why your supplier is also your competitor

This is the part that makes wrappers different from any reseller business I've seen.

Claude Pro at $20 a month includes Claude Code. ChatGPT Plus at $20 includes Codex. GitHub Copilot Pro at $10 includes both Claude Code and Codex agents. If your wrapper does something general (write, summarize, code, chat with files), the vendors will keep folding it into plans your customers already pay for.

So ask yourself the question I'd ask any wrapper founder: if the model vendor shipped your main feature free next quarter, what would your customers still pay you for?

If the answer is "nothing," you don't have a pricing problem. You have a product problem. Chegg's story is the extreme version: read how Chegg lost half its revenue when AI gave its product away.

MOATS ScorecardScore your business on the five moat factors.Open the free tool →

When is a wrapper the wrong business?

I'd walk away, or rethink, if:

Your only edge is a prompt. Someone will copy it, and the vendor will probably build it in.
Your buyers are developers who could call the API themselves for less.
Your best customers are also your most expensive, and you can't price their usage without losing them.
You depend on one model with a scheduled price rise and no fallback.

If two or more of those are true, narrow the job, go vertical, and price against the labor you replace.

Read the whole book free

The Subscription Playbook as a free ebook and audiobook. Australian business owners can claim a free hardcover, shipped free.

Claim the free hardcover Get the ebook

What to do next

Name the job and the customer. "Quotes for trades businesses," not "AI writing."
Cost your typical and heaviest workflow at today's and any announced future prices with the token pricing guide.
Set plans with allowances and marked-up overage. Aim for at least 60% gross margin after inference.
Add a second model for simple tasks and test it on your real work.
Score your moat and pick the one MOATS factor you'll build this year.
Read the Cursor and Lovable stories for the playbook in action: how Cursor broke "unlimited" and how Lovable sells a business that builds businesses.

Then go back to my dog biscuits. A template website and a supplier got you into the market. Only the moat keeps you there.

Frequently asked questions

Are AI wrappers dead?

No, but thin ones are fragile. Lovable said in 2026 it was tracking toward about $600 million ARR, built on third-party models, and SpaceX agreed to buy Cursor in June 2026, according to Wikipedia's SpaceXAI entry. But wrappers whose only edge is a prompt get copied quickly, and model vendors keep adding the same features to their own $20 plans.

Should I let customers bring their own API key?

It can work if you charge for your platform separately. Clay lets customers use their own API keys to skip its Data Credits, but they still pay for Actions, the work Clay's platform does. Bring-your-own-key removes your inference risk, but it also removes your resale margin, so price the workflow, not the tokens.

How much should an AI wrapper charge per month?

Anchor on value, not on ChatGPT's $20. General assistants cost about $5 to $30 a month in 2026, so a general-purpose wrapper struggles above that. Vertical tools priced against labor charge far more: Harvey is reported at $1,200 to $2,000-plus per seat per month. The narrower the job and the more valuable the time saved, the higher you can go.

Which model should an AI wrapper use?

Usually several. Route simple, high-volume tasks to small models costing about $0.10 per million input tokens, and save flagship models for the hard steps. Avoid building only on a model with a scheduled price increase unless your pricing already assumes the higher rate.

Can I build a wrapper with Australian customers and US costs?

Yes, but price in a currency buffer. Model vendors bill in US dollars, so a weaker Australian dollar raises your costs even if usage is flat. Your A$ price also needs to cover the 10% GST you collect from Australian customers. Review both every quarter.

Get both books free

The Subscription Playbook (ebook + audiobook) and Feed a Starving Crowd, instantly. In Australia? Claim the free hardcover.

Robert Coorey, MBA

Robert is the author of The Subscription Playbook and Feed a Starving Crowd and a Director at BPAG Accountants in Sydney. More about Robert.

Sources

Facts and prices checked 27 October 2026. Spotted something out of date? Tell us.

How to Price an AI Wrapper (and Build Something Worth More Than the API Underneath)