How to price credits in your AI app

Most AI builders arrive at credits the same way. The model bill goes up, a few heavy users eat the margin, and a monthly plan suddenly feels risky. So you invent a credit, pick a number that looks like what other apps charge, and hope. Then the questions start: how many credits does a job cost, what happens when it fails, and why did this person sign up, spend their free credits on setup and never come back?

That's a good question, and it has a practical answer. Treat this post as a worksheet: six steps, one worked example with realistic sample numbers, and the rules that make people trust a balance.

What is a credit, really?

A credit is a promise: this many credits get you one finished thing. One summarized report. One product photo. One minute of cleaned audio. People can picture a report. Nobody can picture 14,000 tokens, and they shouldn't have to. Tokens, GPU seconds and API calls are your costs. The credit is what the customer buys.

So step one is naming the unit of value: the smallest result someone would happily pay for on its own. If your app does several jobs, give each one a fixed credit price (a report is 20, a quick rewrite is 2) rather than a price that floats with the length of the prompt. A fixed price lets people decide before they press the button. A floating price makes every click a small gamble.

How are AI credits calculated?

Step two is the arithmetic. You need four numbers: what one attempt at the job costs you, how often an attempt fails, the gross margin you want, and how many credits one job takes. Then three lines:

1. cost per delivered job = cost per attempt ÷ (1 − failure rate)
2. price per job = cost per delivered job ÷ (1 − target margin)
3. price per credit = price per job ÷ credits per job

The first line is the one people skip. If you refund failed jobs (you should, see below), every failure is a cost with no revenue. A 10% failure rate means each finished job carries a ninth of another attempt. Put your support time and payment fees into the cost per attempt too, or lower the margin you aim for to cover them.

So in this example, how much does 1 AI credit cost? About three and a third cents, because you decided a report is 20 credits. There's no market price for a credit. It's whatever your cost, failure rate and margin say, divided by a number you picked. The calculator at the top of this page does the same math; try it with your own numbers before you copy anyone else's price.

AI credits vs subscription: which shape fits?

Step three is choosing how people pay, and the answer comes from how often they need you. Look at how often the job really happens in someone's life or week, not how often you wish they'd open the app.

The freelancer-or-agency question above has a simple answer: offer both, and let rhythm sort them. Packs for people who show up a few times a quarter, a plan for teams who'd rather not think about it. If your product is occasional by nature, don't force a monthly plan and then measure daily use. One builder of an insurance-questions app put it well: the hard part is getting someone back months later, when their policy renews. That's a reminder at the right moment, not a streak (more on that in why users don't come back).

How many free credits should you give?

Step four. An open-ended free tier is a monthly bill you pay for people who may never pay you. Free credits are better because they have an edge, but only if that edge sits after the first finished job. The most common mistake in these threads is a welcome balance that runs out during setup.

These are single makers' own numbers, not benchmarks, but the pattern is clear. Size the welcome balance from the job, not from a round number: enough for the first finished job, plus room to try again if the first one wasn't quite right. In the worked example, 200 credits is ten reports and costs you at most about $1.67 per signup. That's a free tier with a ceiling you know in advance.

Then let people earn more instead of handing out more. Give a few credits for setup steps that make the next job better (connecting a data source, saving a template), and a real bonus for bringing a friend, paid only when the friend finishes their first job. In the example, 500 credits each for a referral is 25 reports, about $4.17 in cost per person, and you only pay it after a new user got value. Our referral guide covers why the friend's first finished job, not the signup, is the moment to reward.

The rules that make people trust a balance

Step five. Credits feel fair when nothing about them is a surprise. These are the rules that come up again and again when people complain about AI credits:

1. 1

   Show the balance and the cost before the job runs

   "This report costs 20 credits. You have 180." On the button, not in the docs. If a job might cost more (a longer file), say the range first.

2. 2

   Refund failed jobs automatically

   If the job errors, times out or returns nothing usable by your own standard, give the credits back without a support ticket. You already priced failures into the cost per job, so this costs nothing extra.

3. 3

   Never silently expire paid credits

   Credits someone bought should stay theirs. Free or earned credits can have an end date, as long as it's stated up front and shown next to the balance.

4. 4

   A receipt for every change

   Every credit in or out gets a line: what happened, how many, the balance after. It answers "where did my credits go?" before anyone has to ask.

On the expiry point, the objection is real. A developer on Hacker News couldn't see why credits they had paid for should have an end date, and the replies, several of them openly guessing, never settled why. If you need a cost ceiling, put it on free credits, not on money people already gave you.

How to design credit packs

Step six. Three packs is enough. One small pack so a new buyer can say yes without thinking, one middle pack most people pick, one big pack for heavy users. Give a bonus on the bigger packs, keep the numbers round, and label each pack in jobs as well as credits, because "35 reports" means more than "700 credits".

Check the last column every time you add a bonus. Bonuses come out of margin, so the biggest pack is where you're thinnest. If it drops below the margin you can live with, shrink the bonus before you raise the small pack's price. And when your model costs fall, you can pass some of that on as a bigger bonus without touching the credit price people already know.

Set it up

You need two separate things: a way to take money, and a way to keep balances honest. The second is where most homemade credit systems go wrong, with a single number in a users table that nobody can explain when it looks off.

Lurk keeps that part for you from day one. You create a Credits unit and a few rules: a welcome balance on signup, a spend when a report runs, a refund when one fails, a bonus for both people when a friend finishes their first report. Your app sends events, Lurk changes the balance, and every change gets a receipt with the balance after it. Lurk doesn't meter your model costs; the cost math above is still yours.

You sell credit packs through your own Stripe account. Bought credits never expire, earned and bought credits are kept apart, and a refund takes back exactly the pack that was bought. Start from the Onboarding and Referral templates for the free credits, then add Coin packs or Top up and earn for the paid ones.

Lurk's own price is a plan plus usage, with a free plan to start on; see pricing. Then paste the prompt below into your coding agent.