How to protect free requests in an AI coding tool
AI coding assistant free tier abuse turns premium requests into bills you can't recover. How to cap and route free requests, and where each fix breaks.
$13,200 of your October model bill went to accounts that were never going to pay you, which is why you're looking at the best chart your company has ever made. New accounts for your coding assistant went from 1,100 in August to 3,300 in October. You've already pasted it into the investor update. Then you click through to see what those accounts actually did.
Run this as a scenario, with numbers you can swap for your own. Your free plan gives each account 50 premium requests a month, and a premium request averages $0.11 because it drags whole files along as context. Of October's 3,300 new accounts, 2,400 spent all 50 inside their first two days and never opened the tool again. None of them upgraded.
Take those 2,400 out and October had 900 new accounts. August had 1,100. The chart in your investor update is going down.
Then you open the bill.
2,400 accounts x 50 requests = 120,000 premium requests
120,000 requests x $0.11 = $13,200Your 310 paying customers bring in $6,200 a month at $20 each, and their own usage, at 90 requests each, costs about $3,069 (310 × 90 × $0.11). So the month's model bill is $16,269 against $6,200 of revenue, and 81% of that bill went to accounts that left.
This is what AI coding assistant free tier abuse looks like from the inside. Your model provider charged for those tokens as they were used, weeks before any of those accounts could have become a customer. A leaked coupon can be voided. A free request has already been answered, so deleting all 2,400 accounts tonight gets you back $0.
The bill also did something worse. You'd set a $15,000 monthly limit on your model account, and it was reached at 2:06 p.m. on the 28th. For the next 3 hours and 20 minutes, every agent run in your product failed, including for the people paying you. One of them wrote: "Your agent died halfway through a migration and left my branch broken. What am I paying for?" Eleven customers asked for their month back, which is $220. Two cancelled. A third stopped opening the tool and hasn't said why.
You had $84,000 in the bank and a burn of $6,000 a month, so 14 months of runway. Add $13,200 and the burn is $19,200, which leaves 4.4 months. The investor update is still open in the other tab, and you now have to choose between cutting the free plan that brought you all 310 customers and paying $13,200 a month to people who were never going to buy.
Why AI coding assistant free tier abuse costs so much
A premium model that writes and fixes code is something people already pay for, so a free allowance has a cash value to anyone who can open more of it. Each request is also expensive to serve, and one agent instruction can mean many model calls. It's the same pattern as credit farming, with an audience that writes scripts for a living.
The allowance also hangs on something that's cheap to make again: an account or an install. So the obvious fix is to count installs. Your extension creates an ID the first time it runs, and you grant the allowance once per ID. That holds until there's another install, because an ID only knows about one copy of the software and new copies are free. It's wrong about honest users too: a real developer has a work laptop, a home desktop and dev containers that get rebuilt every week.
That leaves two jobs: limit what any one account can cost you, then make new accounts harder to collect on.
Sort free requests by what they cost you
Not every feature needs the same limit. Plot each one by what it costs you against how much it helps someone decide to pay.
The top left is your free plan. The top right is worth offering in small, counted amounts. The bottom right is what's most worth scripting, and a new user can decide to buy without it. How many free credits to give starts from this split.
Put a ceiling on every free request
Give free traffic its own model key and budget. Paying customers get a separate key or limit, so free use can't reach their quota. Say that's $150 a day for free traffic. Make it a hard cap, because some budgets only send an alert.
Cap the free-model fallback. A cheaper model with no limit feels close to free. Say the fallback costs $0.005 a request. An account scripted to send 3,000 a month costs $15, nearly three times its whole premium allowance of $5.50 (50 × $0.11). Uncapped, your fallback is an open model endpoint that you pay for. Thirty requests a day is 900 a month, or $4.50.
Hold costly models until the account is verified. An account that never confirms its email should cost you nothing expensive.
Send every free request through one gate, and record what it cost on the account.
Now you can work out the most one account can cost you in a month:
With these numbers, that's (50 × $0.11) + (900 × $0.005) = $5.50 + $4.50 = $10.00. That's also the price of the problem: 1,000 accounts × $10 = $10,000 a month. Every limit here works per account, which is why each one stops short.
The daily budget catches runaway spend, but it can't tell who spent it. If farmed accounts use up the $150 by 1 a.m., your real free users, the ones who might upgrade, see a capacity message for the rest of the day. And someone has to keep resetting that number as you grow.
An account usage cap stops one account overrunning and does nothing about the next, which arrives with a full allowance. It also lands on your most engaged free user halfway through a task.
Rate limits catch loops and bursts. They miss anyone under the limit, and 50 requests over two days is about one an hour. Coding tools have it worse, because a real agent run can fire dozens of model calls in a few minutes. Set a limit for chat and you kill real agent sessions. Set it for agents and it slows nobody down. There's more on what rate limits stop.
Basic account checks and how they break for developers
So you try to slow down new accounts, starting with email. A unique-email rule compares text, and one mailbox can have many spellings. Gmail ignores dots, so jane.doe@gmail.com and janedoe@gmail.com reach the same inbox, and mail to jane+tools@gmail.com lands there too. That's how email works, and developers use it for honest reasons like sorting mail and testing their own apps.
| Basic fix | What it catches | What it misses | Who it wrongly blocks | Cost to keep it working |
|---|---|---|---|---|
| Email verification | Typos and dead addresses | Any real second inbox | Developers behind strict mail filters | Deliverability work |
| Unique emails | The same address twice | Every different address | Almost nobody | Very little |
| Disposable-domain blocks | Known throwaway services | New domains and ordinary mailboxes | People using privacy forwarding | A list someone updates forever |
| CAPTCHA | Simple scripts | People signing up by hand | People using screen readers | Fees and lost signups |
| Phone requirement | Casual second accounts | Anyone with another number | Developers without a reliable mobile number | A charge for every text |
| Card requirement | Casual second accounts | Anyone with another card | Students and people who want to try first | Fewer signups |
| One per IP address | Repeat signups on one connection | Any other network | Offices, bootcamps and campus labs | Exceptions and tickets |
| VPN blocks | Known VPN ranges | Everything off your list | Staff on a company VPN | Another list to maintain |
The last two rows hurt most in this market. Say a 40-person engineering team tries your tool from one office connection. A one-per-IP rule tells 39 of them no, and that was your best sale of the quarter. VPN blocks do the same to anyone whose employer requires one. CAPTCHA has its limits as well, and so does asking for a card.
Two more fixes tempt everyone. Cut the allowance from 50 to 20 and each account costs $2.20 instead of $5.50, but someone after 1,000 requests just needs 50 accounts rather than 20, and every honest developer gets too few requests to finish one real task. Or review new accounts by hand: at one minute each, 2,400 accounts is 40 hours, and by the time you reach an account the model has already been paid.
Stack all of it and a developer who wanted to try your extension has to confirm an email, solve a puzzle, add a phone number and enter a card before their first completion. Your records still can't answer the question the free plan depends on: has this person already had these requests?
Evaluate the claim before the allowance goes out
Every control above limits one account or makes the next one a little harder to open. AbuseZero takes on whether this claim should go through, and we're proud of how. One API call, made before an account gets its premium requests, is all it takes. You also get a proprietary algorithm, with AI that checks the claim again. With multi-layer AI risk analysis, AbuseZero examines 140,000+ signals, detects linked claims and flags abuse that single checks miss. Signals are read together, never one at a time, which is where that table ran out. A domain list goes stale, and AbuseZero checks mailbox aliases, disposable domains and mail routing. An IP rule stops at the 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. In a market where real developers run real agents, it blocks spammy bots, browser automation and malicious scripts while allowing genuine AI agents and assistants. Creating a fresh account isn't enough to erase previous activity.
You get a clear allow or deny, under rules you control. Eligibility can be once per user, once per calendar month or once every 30 days. Observe mode shows you what would be denied before anyone is blocked, and the quickstart covers your first benefit. You don't have to build your own abuse prevention system. Your caps and routing stay yours, because they set what an account can spend. The evaluation decides whether it should get the allowance at all.
Frequently asked questions
How many free requests should an AI coding tool give?
Enough for someone to finish one real task on their own code. Work out your worst case per account first, using the formula above, then calculate what a free user costs you.
Should free users get a fallback model?
Yes, if it's capped. A cheap model keeps people working after the premium requests run out, and a daily limit keeps it from becoming an unlimited endpoint.
Can I get the money back after free requests are farmed?
Not from the requests already answered. You can take back what's unspent and limit the next month.
Keep free requests for people trying your tool
Start with Builder
Pay as you go. Protect your first benefit today.
See pricing
Builder, Pro, Scale and Enterprise.
Talk to us about Enterprise
High volumes and custom terms.