AI credit farming: how fake signups drain your free credits (and how to stop it)

Every AI product with a free tier is giving away something that costs real money: inference. That makes free credits worth stealing. Credit farming is the practice of creating many accounts, by hand or with scripts, to collect the free credits of each one and use them (or resell them) as if they were paid.

This guide explains how farming works, why AI products are hit harder than classic SaaS, and what actually stops it.

Why AI free tiers are a target

On a normal SaaS product, a fake account costs you a database row and maybe an email. On an AI product, a fake account that drains its free tier costs you GPU time or model API spend. Multiply by hundreds of accounts and you have a real bill.

The data backs this up:

• Stripe says its models "detected 6.2x more abusive free trials across Stripe's network" between November 2025 and February 2026, and that AI startups with self-serve signups and direct API access "see 10x more attempted abuse than enterprise AI companies". (Stripe)
• In its first two months, Stripe Radar's trial-abuse control "blocked more than 550,000 free trials with a high risk of abuse", for four AI businesses, "preventing an estimated $4.4 million in downstream losses from compute costs". (Stripe)
• Okta's threat intelligence team saw over 105,000 brute-force signup attempts from 251 distinct IPs against a single AI video service's free trial between June 1 and July 1, 2026. (Okta)

There's a resale market

Farming isn't only hobbyists saving $20. Okta describes a gray market where resellers sell access to models at "5-15% of the official per-token price", sourced from accounts loaded with bonus credits. (Okta) If someone can buy your tokens at a 90% discount, someone else is making money creating accounts on your product.

Smaller teams feel it too. Runbear wrote up how it won $1,000 in AI credits, then saw signups jump to roughly 30 times normal and lost the credits to trial abuse before pausing trials. (Runbear)

The farmer's toolkit

Credit farmers rarely use anything sophisticated. They use whatever is cheapest to repeat:

1. Disposable inboxes. Services that hand out a working inbox for ten minutes. Thousands of these domains exist.
2. Gmail aliases. Gmail ignores dots in the username and anything after a +. One real inbox gives unlimited "different" emails: j.ohn+1@gmail.com, jo.hn+2@gmail.com. See Gmail dots and plus aliases.
3. Fresh IPs. Cloud servers, commercial VPNs, residential proxies or Tor, so IP-based rate limits don't trigger.
4. Scripts. Signup, verify, drain, repeat. Headless browsers and plain HTTP clients both work if your signup is a simple form.
5. The same machine. Even with a new email and IP each time, most farmers run everything from one or a few browsers or devices.

That last point matters: emails and IPs are cheap to rotate; devices and behavior are not.

What doesn't work well

CAPTCHA alone. A CAPTCHA adds friction for every real user and is solved cheaply by solving services. It stops the laziest scripts, not a farmer with a reason.

Blocking disposable domains alone. Necessary, but it doesn't touch Gmail aliases, and new disposable domains appear constantly.

Requiring a card for everyone. It works, but it kills conversion for a free tier whose job is growth. Many teams that went this way, like those discussed on Hacker News, simply ended their free credits. That throws away the growth engine to save the bill.

Strict rules with no feedback loop. geocod.io wrote about choosing to accept some abuse because strict rules blocked real users. (geocod.io) Any defense that can't be tuned will eventually be turned off.

A layered playbook that does work

1. Normalize and deduplicate emails

Lowercase everything, collapse Gmail dots and +tags, and map googlemail.com to gmail.com. Store the normalized form next to the raw one and check for an existing account on the normalized value. This single step defeats the most common trick. Try it on your own list with the email alias checker.

2. Score, don't just block

Combine signals: disposable domain, missing mail server, datacenter or VPN IP, signup bursts per IP and subnet, the same device token across many emails, automation user agents. Each signal alone is weak or noisy. Together they're reliable. Test how this looks with the signup risk test.

3. Gate credits instead of refusing signups

The best response to a suspicious signup is often not a block. Let the account in, but grant a fraction of the free credits until it adds a card, verifies a phone or ages a few days. Real users barely notice. A farmer who needs dozens of accounts loses the payoff. We call this credit gating, and it's the limit decision in BurnGuard.

4. Watch before you enforce

Run detection in monitor mode for a week. Look at what would have been blocked and why. Then enforce, and keep a one-click way to mark false positives as legit.

5. Measure the money

Track what flagged accounts actually consumed. "We blocked 3,000 signups" is abstract; "flagged accounts burned $1,840 of inference last month" gets a cofounder's attention. Estimate your own number with the free-tier abuse calculator.

Where BurnGuard fits

BurnGuard does steps 1 to 5 in one API call at signup: it normalizes the email, scores disposable domains, mail servers, IP type, velocity and device reuse, returns allow, limit, review or block with readable reasons, starts in monitor mode, and shows dollars saved. It's auth-agnostic, with recipes for Next.js, Supabase, Better Auth and Clerk.

You can also build all of this yourself. If you do, the important part isn't the first version; it's keeping the lists, thresholds and device signals current as farmers adapt.

Key takeaways

• Free AI credits are worth farming because they cost you real inference, and there's a resale market for them.
• Farmers rotate emails and IPs cheaply; devices and behavior are harder to rotate.
• Combine signals into a score, then gate credits instead of blocking whenever you can.
• Monitor first, enforce second, and measure the dollars.