Back to Blog

How to calculate the cost of a free user in an AI app

Work out your real cost per free user with one formula, a worked AI app example and a calculator, then see what duplicate accounts do to your payback.

Sanket DattaFounder & CEO

Picture this. It's 11:20 on a Sunday night and you're fixing slide 14 of the deck for Tuesday's investor meeting. It says your cost per free user is $1.35, free-to-paid conversion is 2.0%, and the free plan pays for itself in 3.9 months. You built those numbers yourself from the invoices and the signups table, and every formula adds up.

Then you open last month's new accounts, which you've been meaning to do since summer, and start reading the rows. By midnight you've stopped counting the addresses that are the same person with a dot moved.

Say the true count looks like this. Of 10,000 new free accounts, 3,000 are extra accounts opened by people who already had one. Each of them took all 200 free credits, at $2.05 apiece in model and account costs. That's $6,150 last month and $73,800 a year, spent on accounts that were never going to pay.

Now take that to the bank balance. Say you have $480,000 left and you burn $40,000 a month, which is 12 months of runway. Take out the $6,150 and you burn $33,850, so the same cash lasts about 14.2 months. The leak has been costing you two months of runway, and nobody on the team knew it was there.

Then there's the deck. Your 10,000 new accounts a month is the growth number in the fundraising story, and at most 7,000 of them are new people. Your real conversion is about 2.9%, so the onboarding rebuild that ate your autumn was chasing a rate that only looked low. Without those accounts, your payback would be about 2.1 months. At 3.9, you decided you couldn't spend more on growth, so you held off on the second engineer. At 3.9, you trimmed the free grant in the spring, and real users started running out sooner. Every one of those calls was reasonable on the numbers you had. Meanwhile each paying customer you won was carrying $30.75 of free credits for accounts that would never pay ($67.50 minus $36.75).

It's 12:40 now. On Tuesday you can show a number you know is wrong, or you can tell an investor you don't know your own unit economics.

Every figure in this scenario is illustrative, and all the arithmetic is below so you can swap in your own.

Free-plan cost per paying customer (illustrative)01020304050607080As reportedWithout duplicatesUS dollars
Free-plan cost per paying customer(illustrative)020406080As reportedWithoutduplicatesUS dollars

The cost per free user formula

A free user costs you what they actually use, plus what it takes to have them at all. In plain terms: take the credits you grant, multiply by the share that gets used, multiply by what one credit costs you, then add every other cost that comes with an account.

Cost per free user=(credits granted×share used×cost per credit)+other per-account costs\text{Cost per free user} = (\text{credits granted} \times \text{share used} \times \text{cost per credit}) + \text{other per-account costs}
Credits grantedCredits usedShare usedCredit spendCost per creditCost per free userOther per-account costs
Credits grantedCredits usedShare usedCredit spendCost per creditCost per free userOther per-account costs

Where to find each input

Credits granted is the easy one, because it's on your pricing page. Share used takes a query. Take everyone who signed up in a given month, wait 30 days so they've had time to finish, and divide the credits they used by the credits they were given.

For cost per credit, divide what you paid your model or GPU provider for free-plan usage by the free credits used in that period. Keep paid usage out of it if you can, since paying customers often run bigger models and would push the rate up. Other per-account costs are the small ones people skip: verification emails or texts, welcome jobs, storage, logs. Add up last month's totals and divide by new accounts.

A worked cost per free user example

In the scenario, an AI app gets 10,000 free accounts a month. Each is granted 200 credits, a credit costs $0.01, and other per-account costs are $0.05.

Real usersDuplicatesAll accounts
Accounts7,0003,00010,000
Average credits used100200130
Credit spend per account$1.00$2.00$1.30
Other per-account costs$0.05$0.05$0.05
Cost per account$1.05$2.05$1.35
Monthly spend$7,350$6,150$13,500

Duplicates are 30% of the accounts and about 46% of the spend (6,150 / 13,500). Now add revenue. Say 200 accounts become paying customers at $25 a month, all of them from the real users, and at a 70% gross margin each one brings in $17.50 a month.

MetricAs things standWithout duplicates
Free-to-paid conversion2.0% (200 / 10,000)about 2.9% (200 / 7,000)
Free-plan spend$13,500$7,350
Free-plan cost per paying customer$67.50$36.75
Months to pay that back at $17.50about 3.9about 2.1

Why a blended cost per free user hides the problem

Look at the $1.35 again. No account in that table costs $1.35. Real users cost $1.05, duplicates cost $2.05, and the average sits between them describing nobody.

It also barely reacts when things get worse. Let the duplicate share climb from 30% to 40% on the same 10,000 accounts and duplicate spend jumps from $6,150 to $8,200, while the blended number moves from $1.35 to $1.45. Ten cents doesn't get anyone's attention in a monthly review.

The average stays quiet because every tool you own counts accounts. A new email gets a new row, and a new row looks like a new user to your analytics, your billing system and your signups table. The people behind those extra rows also behave differently. They came for the credits, so they use all of them. That's the pattern behind credit farming, and in an AI app the money's gone the moment a credit is used.

Free user cost calculator

Enter your numbers below. Treat each new free account as one claim, and for the cost of one claim use what a duplicate costs you, which is the full grant plus other per-account costs. With the scenario above you'd enter 10,000, $2.05, 30% and 200, and get $6,150 a month. Everything runs in your browser.

Repeat claims per month
1,200
Monthly cost of repeat claims
$2,400
Yearly cost of repeat claims
$28,800
Reported trial-to-paid rate
5.0%
Rate among first-time claims
5.9%

More on the costs that never reach an invoice is in the real cost of free trial abuse.

Why cutting free credits doesn't fix it

The first move most founders reach for is a smaller grant. Say you halve it to 100 credits. Those 3,000 duplicates took 600,000 credits between them, so if the same people still want 600,000 credits, they now need 6,000 accounts. At $1.05 each that's $6,300, which is $150 more than before. Your signup chart looks better, and the real users who needed more than 100 credits to see what your product does have hit a wall. A usage cap has the same limit: it holds per account, and a new account comes with a new cap.

How to estimate your duplicate share with basics

You can get a rough number this week with nothing but your own database.

Count unique emails. Take last month's new accounts and count each inbox once, so a Gmail address written with dots or a plus tag isn't counted twice. The gap between accounts and unique inboxes is your first estimate.

Review a random sample. Pick 100 new accounts at random. Have one person open each and mark it yes, no or can't tell: is this a second account for someone who already has one? Count only the yeses, and each is one percentage point. At three minutes an account, that's five hours. It's also rough. If 30 come back yes, the true share is likely somewhere between about 21% and 39%.

Take the higher of the two. That's your floor. Put it in the calculator and treat the result as the least you're losing.

It's a floor because both methods only find careless duplicates. An email count can merge addresses that are visibly the same inbox, and that's all it can do. Someone on an unrelated address looks new to the query and new to your reviewer. As an illustration, say your count finds 800 of the 3,000 duplicates in the scenario. You'd report 8% and relax, and the other 2,200 would go on costing $4,510 a month.

What the other basics catch, and how they break

BasicWhat it catchesWhat it missesWhat it costs to keep working
Unique-email countThe same address used twiceAny new addressA report someone has to rerun
Email verificationTypos and addresses that can't receive mailAnyone with a second real inboxA step every real user has to finish
Disposable-domain blockThrowaway services on your listNew domains and ordinary mailboxesA list someone updates forever, and real customers on odd domains turned away
Gmail dots and plus addressingAliases of one inboxEvery address that isn't an aliasUpkeep as mail services change, and merged accounts of people who sort their mail
Usage capRunaway spend on one accountThe next accountSupport tickets from real users at the limit
Manual review of a sampleWhat a careful person can spotWhatever looks ordinaryHours each month, and it stops nothing

These break quietly, too. A blocklist doesn't know about a domain registered yesterday. The Gmail rule handles Gmail, and then the next provider needs its own. The sample goes stale in a month, and all it ever told you was how much you'd already lost. Add a CAPTCHA and you've filtered some bots while the person opening a fifth account by hand carries on.

Replace your floor with a measured number

You don't have to build your own abuse prevention system. Those basics look at one email address at a time. AbuseZero's multi-layer AI risk analysis examines 140,000+ signals, reads them together, never one at a time, detects linked claims and flags abuse that single checks miss. You also get a proprietary algorithm, with AI that checks the claim again. AbuseZero recognizes repeat devices across new accounts, even when email addresses keep changing, and detects repeated payment fingerprints across users, even when new accounts and emails hide the connection. Creating a fresh account isn't enough to erase previous activity. You make one API call, you get a clear allow or deny, and the rules stay yours.

For this post, Observe mode is the part that matters. It lets you see what would be denied before anyone is blocked, so you can compare your floor with what happens on your own traffic. See protect and observe and test changes safely. Evaluations in Observe mode are real evaluations and use credits. There's more in shadow mode for abuse rules.

Frequently asked questions

What's a good cost per free user?

There's no single good number, because it depends on what a paying customer brings in. Divide your free-plan spend by new paying customers, then divide that by monthly gross margin per customer. That's your payback in months, and it's the figure to judge.

Should duplicates count in my cost per free user?

Count their spend, because you paid it. Keep them out of the user count, and report real users and duplicates on separate lines so the average can't hide either one.

How often should I recalculate it?

Monthly, by signup month, and again whenever you change the grant or your model. A new model price moves cost per credit straight away.

Can I measure duplicates without blocking anyone?

Yes. The email count and sample review give you a floor, and Observe mode shows what would be denied while every claim still goes through.

Put a number you trust in the deck

Keep reading

Copying your prompt…

Copied the to your clipboard. Give it to your coding agent & it will handle the rest.