Back to Blog

How to protect the free tier of an AI agent platform

One free agent run can trigger dozens of model calls. What that costs, which limits help and which basic signup fixes keep breaking.

"Our lead routing agent has been failing since early this morning. It got through 212 of our 340 new leads, stopped on a step that says 'capacity unavailable', and now I can't tell which ones were assigned. Two of my reps have already called the same person. We pay you $299 a month so we don't have to babysit this. Should I switch it off and do it by hand? I need to know before standup."

That message is a scenario, and so is the rest of this morning. The prices below are assumptions to swap for your own.

You run a hosted agent platform. People describe a goal, connect their tools, and your product acts until the job's done. It's 7:48 a.m. on the 14th, and your model provider's dashboard says the $9,000 monthly limit was reached at 6:37 a.m. For 71 minutes, every agent on your platform has been getting a quota error. Agents don't stop cleanly, either. That lead routing agent had already changed 212 records, and your customer is now working out which ones.

The logs show where the money went. Your 36 paying customers and your honest free users spend about $180 a day, or $2,340 over 13 days. Nearly everything else went to 141 free accounts that were opened, drained and left. This is AI agent platform free tier abuse, and it's priced differently from most free tiers, because yours gives away 20 runs a month and nothing says how big a run is.

Assume $3 per million input tokens and $15 per million output tokens. Say each agent starts with 6,000 tokens of context, adds 2,000 more per step and writes 300 tokens per step.

5-step run:    50,000 input + 1,500 output tokens = $0.1725
25-step run:  750,000 input + 7,500 output tokens = $2.3625

As you priced it:  20 runs x $0.1725 =  $3.45 a month
Fully drained:     20 runs x $2.3625 = $47.25 a month

141 accounts x $47.25             = $6,662.25
141 accounts x 20 runs x 25 steps = 70,500 model calls

Five times the steps cost nearly 14 times as much, because every model call resends everything the agent has seen so far.

Model cost of one run, illustrative00.511.522.5510152025US dollarsSteps in the run
Model cost of one run, illustrative012510152025US dollarsSteps in the run

Those 36 customers at $299 bring in $10,764 a month. A normal model bill is $180 × 30 = $5,400, which leaves $5,364 for everything else. With the 141 accounts, the bill is $12,062.25, and that's $1,298.25 more than your entire revenue. Say six customers ask for this month back: 6 × $299 = $1,794. Say your other costs are $9,000 a month, so you normally burn $9,000 minus $5,364, or $3,636. The farmed runs plus the refunds come to $8,456.25, or 2.3 months of that burn, gone in under two weeks.

None of it comes back. Your provider billed those tokens as they were processed, so deleting all 141 accounts tidies your user count and leaves the invoice where it was. That invoice arrives on the provider's schedule, while your customers renew one at a time through the month.

You still have to answer that message. Raising the limit brings your customers back in a minute, and the free accounts too, because nothing has stopped them. Holding it keeps 36 paying customers down while you hunt. You raise it from your phone, and then you get to the question you didn't want this week: should you keep a free tier at all, when it found you most of those 36 customers?

The next part really did happen.

From our early beta

Fake users drained one team's shared model provider quota. Their paying customers couldn't use the product until the team raised its budget to cover the abuse.

When fake users arrive through a free tier, that's the capacity cost of repeat claims.

Why AI agent free tiers have no natural ceiling

Most free tiers count something with a roughly fixed price. An image is an image. A run isn't like that, because an agent keeps going until it decides it's done or you stop it. The goal, the tools and the pages they pull in set the cost, and the person choosing those isn't paying.

loop[Up to 25 steps]Start one runWhole history so farNext actionBrowse, search or fetchResult joins the historyFinished workYour meter counts 1Your provider bills 25 callsFree accountYour platformModel providerToolsFree accountYour platformModel providerTools
loop[Up to 25 steps]Start onerunWholehistoryso farNextactionBrowse, search orfetchResult joins thehistoryFinishedworkYourmetercounts1Your providerbills 25 callsFreeaccountYourplatformModelproviderToolsFreeaccountYourplatformModelproviderTools

Tools usually bill on top of that. Prompt caching, where offered, lowers the input side, but the cost still climbs with every step.

A run is also worth having. 25 steps of browsing and writing is work someone could sell to a client, and a new account has 20 more runs. That's credit farming with a much bigger payout per credit.

The obvious fix is a lower step cap, and it does help. At 10 steps a drained account costs 20 × $0.495 = $9.90 instead of $47.25. But free tier cost is accounts × runs per account × cost per run, and a step cap only shrinks the last term. The same $6,662.25 now takes 673 accounts rather than 141 ($6,662.25 ÷ $9.90), and accounts are free to open. Meanwhile an honest builder whose agent needed 14 steps gets a failed run at step 10, in the first session they've ever spent with you, and you still paid $0.495 for that failure.

Free run limits to set first

These won't stop a new account. They do cap what any one account can cost.

Meter steps or dollars, not runs. Turn "20 runs" into "100 steps". Someone running five-step agents still gets 20 runs. A 25-step run now uses a quarter of the allowance, so a drained account costs 4 × $2.3625 = $9.45 instead of $47.25.

Give each free run a dollar budget, enforced on your server before every model call. When it runs out, return what the agent has so far.

Keep free traffic off the paying quota. If your provider lets you set a separate limit per key or project, give the free tier its own. A bad day then ends the free runs and your customers never see it. Farmed accounts can still empty that budget, though, and your honest free users get the quota error instead.

Run paid agents first. Free runs wait in their own queue, so they can't delay a customer's job.

Start nothing before the email is confirmed, and hold the costliest tools back until someone has paid.

The basic signup checks, and how they break

Those limits work per account, so next you'll want to know whether a new account belongs to someone who's already had their runs. Most teams reach for these, and each one catches something.

Basic fixWhat it catchesWhat it missesWho it wrongly blocksWhat it costs to keep working
Email verification, unique emailsDead addresses, exact repeatsAny real second inboxBuilders whose code lands in spamDeliverability work
Disposable-domain blocksKnown throwaway servicesNew domains, ordinary mailboxesPeople using privacy forwardingA list someone updates forever
Gmail dots and plus addressingOne Gmail inbox written several waysA second real mailboxPeople who sort their mail with plus addressesUpkeep as mail services change
CAPTCHASimple signup scriptsPeople signing up by handPeople using screen readersFees and lost signups
Phone or card requirementCasual second accountsAnyone with another number or cardStudents and weekend buildersText fees, fewer first agents built
One per IP, VPN blocksRepeats on one connection, known VPN rangesAny other networkCoworking spaces, engineers on company VPNsExceptions and another list
Rate limitsToo many runs started at once141 accounts under the limitTeams starting runs in batchesTuning per endpoint
Manual reviewWhat a person spots in a run logWhatever ran before anyone lookedWhoever looks odd to a tired reviewerHours every month

A rate limiter sees one request to start a run. The 25 model calls after it come from your own servers, so the limiter never counts them. Manual review runs into the clock: at two minutes each, 141 accounts take 282 minutes, and the money was spent before you opened the first one. CAPTCHA and VPN blocks carry costs of their own.

Gmail ignores dots, and plus-tagged mail reaches the same inbox, so one mailbox can look like many addresses to a rule that compares text. Stack every row above and your signup form asks for a verified email, a puzzle, a phone number and a card, and the lists still need updating next week. What's left is the moment the free runs are about to be granted, and one question about that claim: should it go through?

Evaluate each claim before the free runs go out

AbuseZero was built for that moment. Your server makes one API call and gets a clear allow or deny, under rules you control, so you don't have to build your own abuse prevention system or keep a domain list alive. Where each row above looks at one thing, multi-layer AI risk analysis examines 140,000+ signals, detects linked claims and flags abuse that single checks miss. There's a proprietary algorithm too, with AI that checks the claim again. AbuseZero recognizes repeat devices across new accounts, even when email addresses keep changing, and sees beyond an IP address with VPN, hosting, network reputation, Tor and proxy intelligence. It blocks spammy bots, browser automation and malicious scripts while allowing genuine AI agents and assistants, which matters on a platform where your best users build those for a living. And creating a fresh account isn't enough to erase previous activity, which is the gap every cap above leaves open.

Eligibility can be once per user, once per calendar month or once every 30 days, so signup runs and a monthly refresh can each have the right limit. Observe mode shows you what would be denied before anyone is blocked, and the quickstart walks you through your first benefit.

Frequently asked questions

How many steps should a free agent run get?

Enough to finish the task your onboarding shows. Measure the steps your first-run template takes and add a margin. Work out what a free user costs before you pick the number.

Should free runs use a cheaper model?

Where it can still do the job, yes. At a fifth of the assumed price, $47.25 becomes $9.45. A weaker model may take more steps, though.

Do I still need caps if I evaluate claims?

Yes. AbuseZero evaluates a claim when it's made. Your application still runs the agent, meters its steps and enforces usage limits.

Is every second account on an agent platform abuse?

No. Builders keep work and personal accounts, and some open one just for testing. Write down what each person gets before you enforce anything.

Keep your free runs for people building agents

Keep reading

Copying your prompt…

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