Budget the free user like CAC
Work backwards from conversion. If roughly 3 in 100 free users convert to a $29 plan with healthy margins, then spending $0.20 a month per free user on model costs is cheap acquisition, and $5 is not. Pick the number deliberately, then engineer to it: quota times cost per action must land under budget for the heavy free user, not the average one. Cost per action comes from the arithmetic in estimating AI feature cost; on a budget model, a typical transform action costs about $0.003, so even a generous 50 action monthly quota budgets around $0.15 per fully active free user. The math is forgiving. The tail is not.
The three layer limit design
- Visible quota: N actions per month, framed in product terms (summaries, messages, generations), reset monthly, upgrade prompt when reached. This is the layer users see and understand.
- Invisible guards: input length bounds, tight
max_tokens, and per user request rate limits at your proxy, the mechanics from rate limiting AI endpoints per user. These bound the worst case cost of any single action so the quota math holds. - The backstop cap: a hard monthly spending cap per user in dollars, enforced before the call. Quotas count actions; caps count money, and money is what you are protecting. When the two disagree (an action turned out expensive), the cap wins.
Put free users on a budget model and use downgrade as the soft landing: a free user at their cap gets gpt-5.4-nano instead of a wall, per model downgrade on budget. Paid tiers get the better model, which makes the upgrade tangible.
Abuse: what actually happens and what catches it
Three patterns account for most free tier bleeding. Signup farming: scripted accounts each consuming the free quota, defended by normal signup friction plus the observation that farmed accounts have machine usage patterns. Bot driven accounts: one account, inhuman volume, caught by velocity, no human sends 400 requests in ten minutes. And quota resale: your free tier wired into someone else’s product as a backend, visible as sustained, regular, round the clock usage. All three share a signature, volume and regularity no human produces, which is why per user velocity detection, the same machinery as AI agent loop detection, is the practical abuse alarm: it fires on the pattern within minutes, not on the invoice within weeks.
Weckr covers this layer end to end for free tiers: per plan spending caps (a free plan with a $0.50 monthly cap is one setting), block or downgrade enforcement before the call, and velocity alerts to Slack when any account burns abnormally fast. The free tier stays an acquisition tool with a known price tag.
FAQ
Should my AI product have a free tier?
Usually yes, free tiers remain the strongest acquisition tool for developer and prosumer products, but an AI free tier is different from a normal one: every free action has a hard marginal cost. The design question is not whether to offer free AI, it is how to bound the cost per free user so conversion economics stay positive. Unbounded free AI is a donation program with signup friction.
How much should a free tier cost me per user?
Decide it explicitly as your acquisition cost budget. If free to paid conversion is around 3 percent and your paid plan margins are healthy, a free user costing you $0.10 to $0.50 a month in model spend is a reasonable customer acquisition cost; $5 a month per free user usually is not. Pick the number first, then set the quota and model so the p95 free user stays under it.
How do free AI tiers get abused?
Predictably: scripted signups farming free quota across many accounts, one account driven by a bot loop instead of a human, and free tiers of writing or coding tools getting wired into other products as a free backend. The common thread is volume no human produces, which is why velocity detection, calls per user per minute, catches most abuse before the invoice does.
Should free users get a worse model?
Usually yes, and it is standard practice: free tier on a budget model, paid tiers on better ones. It bounds free cost, creates a real upgrade incentive, and current budget models are good enough that the free experience stays convincing. The downgrade path also works as a cap action: instead of cutting a free user off at the quota, drop them to a cheaper model.
What limits should a free AI tier have?
Three layers: a visible product quota (N summaries or messages per month, chosen from your cost budget), invisible guards (input length and max_tokens bounds, per user rate limits), and a hard per user monthly spending cap as the backstop that blocks or downgrades whatever slips past the first two. The cap is what turns a worst case from an incident into a non event.
Keep reading
Free for them. Bounded for you.
The free tiers that work are the ones where the founder knows exactly what a free user can cost, because the ceiling is enforced, not hoped. Weckr gives every plan a cap and every user a meter, from two lines, free for 50,000 requests a month, which covers a lot of free tier. See caps working on the live demo, or start with the AI cost and margin guide.