How to offer free LLM tokens without funding resellers
Free LLM token abuse is cash you've already spent. What token caps and limits stop, why new accounts walk past them, and how to set a budget you can lose.
"Why did we pay more for tokens last month than customers paid us?"
Your co-founder sent that at 8:37 a.m. on the 3rd, under a screenshot of your model provider's invoice. You've started a reply twice and deleted it, because you don't know.
This is a scenario, so swap in your own numbers, but follow the arithmetic. You've got 190 customers at $40 a month, which is $7,600. The invoice says $8,340. Your free plan gives every account 500,000 tokens a month, you pay a blended $6 per million tokens, and you've always told yourself that free LLM token abuse tops out at $3 an account, because that's what the cap is for.
By 10:45 you've exported usage per account, and there it is:
Paying customers: 400M tokens at $6/M = $2,400
Ordinary free users: 240M tokens at $6/M = $1,440
1,500 new accounts: 750M tokens at $6/M = $4,500
Total: 1,390M tokens at $6/M = $8,340All 1,500 were opened in the last few weeks. Each one used its full 500,000 tokens and stopped. None of them broke a limit, so your cap did exactly what you built it to do, 1,500 times.
Take those accounts out and the bill is $3,840, which leaves you $3,760 of last month's revenue. Leave them in and you're $740 short before you've paid for hosting or anyone's salary. You can delete all 1,500 this morning and the invoice won't change by a cent, because your provider already ran those tokens. With a free plan, the model bill always arrives before the revenue does. You accept that for someone who might upgrade in three months. These accounts never entered a card, so no revenue is coming behind this bill.
Say you've got $84,000 and you spend $7,000 a month more than you earn, which is 12 months of runway. Add $4,500 a month and it's $84,000 ÷ $11,500, about 7.3 months. Nearly five months of your company went to accounts that never sent you a word.
You'd also set an $8,000 spending limit at your provider. It tripped at 3:41 p.m. on the 29th, and every request in your product failed until you raised it at 4:59. For 78 minutes your paying customers got errors because accounts paying nothing had used up the budget. Six of those customers asked for the month back, another $240, so you're $980 short.
Your co-founder has followed up: "Should we just turn free off?" Free is where most of those 190 customers came from. And it's the 3rd, so every one of those allowances came back two days ago.
Why free LLM token abuse gets past token limits
Free tokens are closer to cash than almost anything else you give away. A free seat in a project tool is only worth something inside that tool. A model response can be copied out and used anywhere, and you've paid for it before anyone reads it.
Resale is the risk that makes this worse, and we're describing a risk here, not reporting a case. Anyone holding those 750 million tokens and selling them at half your price would take in $2,250 a month (750 × $3), with nothing to pay for the supply. You may never see proof of that in your logs, and your invoice is the same either way.
The obvious fix is a tighter cap, so say you cut free from 500,000 tokens to 100,000. Your cost per account falls from $3 to $0.60, and that's real. But 750 million ÷ 100,000 is 7,500 accounts, so all you've changed is how many signups the same tokens take. Every honest free user just lost 80% of what you promised them.
The sum behind this has three parts:
Every token limit you can build shrinks the second part or the third. None of them touches the first.
How to set free LLM token limits
Build these anyway. They put a ceiling on what one account can cost, and the last one protects your paying customers. Work out what a free user costs you first, then decide how many tokens to give.
Count tokens on your server, per account. Check the balance before the model call and deduct real usage after it. This stops one account running past its allowance. The upkeep is in the edge cases, because a retry or a dropped stream still costs you tokens.
Cap each request. Set a maximum input size and a maximum output for free users. That stops one huge prompt from eating the allowance, and it means someone pasting a long document gets a cut-off answer.
Add a daily limit under the monthly one. It slows any one account without lowering the total, and it lands on the person trying your product hard on day one.
Put free traffic on a cheaper model. This lowers your price per million. The cost is that free users judge your product by a weaker model than the one you sell.
Keep free tokens inside your product where you can. An API key is the easiest kind of access to pass along. If API access is what you sell, keep free keys low and limit them per developer. You'll lose some developers who wanted to test from code.
Give free usage its own hard budget. Use a separate key or project with its own limit if your provider offers one, and make it a real spending cap that stops requests. Set it to a number you can afford to lose. When it runs out, free stops and paid keeps going.
That last step would have turned your 78 minutes of errors into a problem for free users only. It doesn't bring your free plan back, though. Say the free budget is $1,800 a month. Ordinary and farmed use together run at $5,940 a month, or $198 a day, so the budget lasts about 9 days. It's gone on the 10th, and your honest free users spend the next 20 days looking at "come back next month" because someone else got there first.
Placement is our judgment, not a measurement. Notice the empty corner.
The signup basics, and how they break
Limits can't reach the number of accounts, so the next place to try is the signup form. Each of these does catch something.
Email checks. Verification catches typos and dead addresses, and a unique-email rule catches the same address twice. Neither does anything about a second real inbox, and a code that lands in spam loses you a real person. One inbox can also have many spellings: Gmail ignores dots in personal addresses, and mail with a plus sign and a word after the name still lands in the same place. That's just how email works.
Disposable-domain blocks. They catch known throwaway services and miss any domain that isn't on your list yet. They also turn away people who use privacy forwarding, and someone has to keep that list updated forever.
CAPTCHA. It filters simple scripts, and anyone signing up by hand passes. People on screen readers or privacy tools have a harder time, and you pay for it in fees and lost signups.
Phone or card requirements. Both cut casual second accounts, and both stop working on anyone with another number or card. You pay for every text. Students, people without a reliable mobile number and anyone who wanted to try before paying leave without a word.
One per IP and VPN blocks. These catch repeat signups on one connection and known VPN ranges, and they miss every other network. Offices, campuses and households share addresses, so you can turn away a paying customer's colleague. Staff on company VPNs get caught too, and you've got another list to maintain.
Rate limits. They stop bursts from one source. Spread 1,500 accounts over three weeks and that's about 3 an hour (1,500 ÷ 504 hours), which no sensible limit fires on, while a real office in a busy hour might trip it.
Shorter free windows. Seven days instead of a month limits slow, steady use. An allowance can be spent in an hour, though, and honest users get less time to see the point.
Manual review. It catches what a person can spot. At one minute per account, 1,500 accounts is 25 hours a month, and they don't look wrong: each one is verified and under its cap. Make one tired call and you've banned someone who used all their tokens because they liked your product.
If you're doing this yourself, build the six limits, hold free tokens until an email is confirmed so unverified signups cost you nothing, and add the signup steps your honest users can stand. You'll know your worst month to the dollar. You'll also have lists and thresholds to maintain, and when one quietly stops working, you'll hear about it from an invoice.
Evaluating the claim before the tokens go out
You don't have to build your own abuse prevention system to answer the question all of that leaves open, which is whether this new account should get these tokens at all. AbuseZero exists for that moment, and with it, creating a fresh account isn't enough to erase previous activity. That's the exact gap your 500,000-token cap left. With multi-layer AI risk analysis, AbuseZero 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. Those signals are read together, never one at a time, so you're no longer depending on the one domain list or IP block that gave out above. AbuseZero recognizes repeat devices across new accounts, even when email addresses keep changing. It checks mailbox aliases, disposable domains and mail routing, and it sees beyond an IP address with VPN, hosting, network reputation, Tor and proxy intelligence.
Your server makes one API call before it grants tokens and gets a clear allow or deny, under rules you control. Eligibility can follow the way your tokens work, so a signup grant can be once per user and a monthly allowance once per calendar month. Observe mode shows you what would be denied before anyone is blocked, and the quickstart protects your first benefit.
Keep your caps and your free budget. They limit what one account can cost. The evaluation decides whether it gets the benefit at all.
Frequently asked questions
Do token limits stop free LLM token abuse?
They limit what each account can cost. They don't limit how many accounts there are, so the total still grows with every new signup.
Can I tell if my free tokens are being resold?
Usually not with any confidence. Your logs show what was requested, not who ended up with the output. Plan your limits as if it could happen.
Should I stop offering free tokens?
Treat that as a last resort, because free is often where your paying customers come from. Work through the decision before you remove it.
Give free tokens to people who might buy
Start with Builder
Pay as you go. Protect your free tokens today.
See pricing
Builder, Pro, Scale and Enterprise.
Talk to us about Enterprise
High volumes and custom terms.