How many free credits should an AI app give new users?
Work out how many free credits to give new users. Price the first useful result, add a margin, then see what repeat claims do to the number.
Your cancel form has a box asking what went wrong, and a customer who paid you $49 a month for 11 months just left it empty.
You open their account, looking for anything. Their last session was on the 25th. Between 6:03 and 6:29 a.m. they tried to generate a slide deck 14 times and got 14 errors, and they never came back. They didn't send a ticket or a complaint. You'd have preferred the complaint.
That morning, your model provider's $4,000 monthly limit ran out at 6:03 a.m., and you raised it from your phone at 7:21. For those 78 minutes every paying customer saw the same error. And when you trace where the $4,000 went, you end up at a number you set months ago and never had a way to test: how many free credits to give each new account.
It's a made-up example, so change the figures to fit your product. Your app turns a brief into a deck, a credit costs you $0.01 in model spend, and every new account gets 500 free credits. You pull the month's usage:
- 640 accounts spent all 500 credits within an hour of signing up, then never came back: 640 × 500 × $0.01 = $3,200
- 400 real free users spent 200 credits each on average: 400 × 200 × $0.01 = $800
- 62 paying customers used $820 between them
That's $4,820 for the month, about $161 a day, which is how a $4,000 limit runs out on the 25th. Those same 62 customers pay you 62 × $49 = $3,038. So 640 accounts that have never paid you anything cost more than everyone who does pay you, combined.
The model bill doesn't wait. Your provider charges for tokens as they're generated, and the revenue they were supposed to lead to comes later, or never. You can delete all 640 accounts this afternoon and the $3,200 stays spent, because a used credit is cash that has already left.
The bars are what 640 repeat accounts cost at each grant size. The line is your $3,038 of monthly revenue.
Say your company spends $9,000 a month, model bill included, and has $48,000 in the bank. After revenue you're burning $5,962 a month, which is about 8 months of runway. Without that $3,200 you'd burn $2,762 a month, which is about 17 months. Those accounts are costing you about 9 months of runway.
That customer was worth $49 × 12 = $588 a year. Two more cancel before Friday, so 78 minutes has now cost you 3 × $588 = $1,764 a year, on top of the bill that caused it. The 1st is coming and you have to choose. Cut the free credits and the next real user may not get far enough to care. Keep them and you pay this again.
Why free credit grants go wrong
You sized the grant thinking about one person trying your app once, and nothing shows you the cost until the month is over. What you agreed to pay is the grant times however many accounts get opened, and nothing on your pricing page sets that second number. That's credit farming, and AI apps get hit first because a credit turns into a real cost within seconds.
The obvious fix is cutting the grant, and it does help. At 100 credits, the same 640 accounts cost $640 instead of $3,200. But someone who wants the same 320,000 credits (640 × 500) can open 3,200 accounts instead (320,000 ÷ 100). The honest user, who'll only ever open one account, takes the full cut. If a finished deck takes 60 credits and most people need two tries before they keep one, 100 credits runs out partway through the second. You've paid for a first try and lost the reason anyone would pay you.
So you have two numbers to set. The size of the grant is something you can calculate. How many times one person receives it is something you have to enforce.
How many free credits to give, worked out
Free credits have one job: getting a new person to the first result they'd pay for. Give less and you've spent the money for nothing.
Name the first useful result. Pick the thing someone would be annoyed to lose: an exported deck, a finished transcript, an image they download. A first draft usually isn't it.
Count the credits and the tries. Look at the people who went on to pay, and add up what they spent before their first export. Take the middle value, because a few heavy users drag the average up. Here, that's 60 credits, with 2 results generated before one was kept.
Add a margin. It covers longer inputs, failed runs and showing a colleague. 25% is a fair place to start, so 60 × 2 × 1.25 = 150 credits.
Use your real cost per credit. Divide last month's provider invoice by credits used, so failed and retried calls are counted. Here's the full calculation. At $0.01, 150 credits costs at most $1.50 per account.
Check it earns itself back. Say 4% of 400 real signups pay, which is 16 new customers a month. Each paying customer costs you about $13 in model spend ($820 ÷ 62), leaving about $36 of every $49. If every real signup spends all 150 credits, that's 400 × $1.50 = $600, or $37.50 per new customer, earned back in about a month.
What repeat claims do to your grant
Now put the 640 accounts back in. The grant is still 150 credits, each claim can still cost $1.50, and you still get 16 new customers.
| Who claims the 150 credits | Claims | Cost of grants | Cost per new customer | Months to earn it back |
|---|---|---|---|---|
| Real users only | 400 | $600 | $37.50 | About 1 |
| Real users plus 640 repeat accounts | 1,040 | $1,560 | $97.50 | About 2.7 |
| Real users plus enough accounts for 320,000 credits | 2,534 | $3,801 | $237.56 | About 6.6 |
The last row is the same 320,000 credits collected 150 at a time: 2,134 accounts, rounded up, plus your 400 real users. Only the number of claims changed between those rows, and your well-priced grant now costs 2.6 times as much per customer on the second row and 6.3 times on the third.
The basic limits on repeat claims, and how they break
Most teams try the same list. Each item catches something, but only so far.
Email rules
Verification catches typos and dead addresses, but it only proves someone can open an inbox. Any real second mailbox passes, and real people lose codes to spam. A unique-email rule catches the same address twice and almost never blocks anyone, but one mailbox can have many spellings. Personal Gmail ignores dots, and mail sent to a plus address lands in the same inbox. That's how email works. A disposable-domain block catches known throwaway services, misses newer ones, turns away people using privacy forwarding, and needs someone updating the list forever.
Steps on the signup form
CAPTCHA stops simple scripts. It doesn't stop someone signing up by hand, it gets in the way of people using screen readers and privacy tools, and it costs you signups. A phone or card requirement stops casual second accounts, but not anyone with another number or card. Phone shuts out people without a reliable mobile, and you pay for every text. A card shuts out students and people who want to try first. Weigh that against what you're protecting: a $1.50 grant, guarded by a step that can turn away someone worth $36 a month.
Network limits
One signup per IP address catches repeats on one connection and nothing from any other network. It lands hardest on offices, campuses and households, which is awkward for an AI tool that spreads when one colleague shows another. VPN blocks catch known ranges only, turn away privacy-minded people and staff on company VPNs, and add another list to maintain.
Usage limits
Rate limits catch bursts from one account and can trip on real users in a busy hour, but 640 accounts under the limit look fine. A per-account cap is what your grant already is, and every new account gets a fresh one. The limit worth having is a hard budget for free usage, kept apart from paid usage, so a bad month stops at the free plan. It caps what claims cost without reducing them, and real free users get turned away when it trips.
Manual review
A person can spot things a rule can't, and can also turn away whoever looks odd late on a long day. At one minute per account, 640 accounts is nearly 11 hours a month, and those accounts were empty within the hour. Review can tell you which accounts burned the credits, and by then you've paid.
Worth doing this week
Hold the margin back until someone keeps a first result, and re-run the calculation when your model price changes.
Stack every one of these and your 150-credit grant is still only priced right while claims stay close to one per person.
Keeping claims close to the number you priced
The formula gets the size right and assumes the count. AbuseZero is built to evaluate whether each claim looks like a repeat. Your unique-email rule stops at the text of an address, while AbuseZero checks mailbox aliases, disposable domains and mail routing. Your IP limit knows one address, and AbuseZero sees beyond it with VPN, hosting, network reputation, Tor and proxy intelligence. It recognizes repeat devices across new accounts, even when email addresses keep changing, so creating a fresh account isn't enough to erase previous activity. Its multi-layer AI risk analysis examines 140,000+ signals, detects linked claims and flags abuse that single checks miss. There's also a proprietary algorithm, with AI that checks the claim again.
All of that reaches you as one API call before the credits move, and a clear allow or deny under rules you control. You don't have to build your own abuse prevention system, or shrink a grant you've just priced properly. Each grant can be its own benefit with its own eligibility, Observe mode shows you what would be denied before anyone is blocked, and the quickstart walks you through your first one.
Frequently asked questions
How many free credits should a new user get?
Enough for one useful result, tried about twice, plus a margin of around 25%. Use your own usage logs, because another app's number won't fit.
Should free credits expire?
Usually, yes. Expiry stops unused grants from sitting around as a future cost. It won't slow down anyone opening new accounts.
Can I take free credits back after abuse?
Unused ones, yes. Spent ones are gone, along with what you paid for them, so taking credits back only helps with what's left. That's why AbuseZero evaluates each claim before the grant goes out.
Give your free credits to the people you priced them for
Start with Builder
Pay as you go. Protect your first grant today.
See pricing
Builder, Pro, Scale and Enterprise.
Talk to us about Enterprise
High-volume credits and custom terms.