Why do free tiers no longer work the way they used to? For 2 decades a free user cost some storage and some support queue, so free rounded to zero and an unlimited tier was a cheap way to buy a funnel. Once every meaningful feature acquired a model call, every free action acquired a real marginal cost, and an agentic workflow can spend a meaningful amount before any limit fires, because the cost accrues inside the run rather than at the end of the month.
Free survives, because its job survives. What ended is leaving it open like a faucet. The version that works gates by value rather than time, keeps the cheap deterministic features free and habit-forming, and spends the inference budget on the first session against the user's own data, with limits enforced in the request path rather than discovered in a monthly report.
For 2 decades, free was a marketing expense that rounded to zero. A free user cost some storage and some support queue, so unlimited free tiers were a cheap way to buy a funnel. Then every meaningful feature acquired a model call, and every free action acquired a real marginal cost. Free survives, because its job survives. What ended is the era of leaving it open like a faucet.
What changed in the math?
A classic SaaS free user consumed pennies a month. An AI product's free user consumes inference on every session, and an agentic workflow can consume a meaningful amount before any limit fires, because the cost accrues inside the run, not at the end of the month. Unlimited free at scale is now a decision to fund strangers' compute, and finance will eventually ask who made it.
What the market converged on
Look across AI products and the pattern is consistent: the unlimited tier gave way to credits, daily caps, and metered allowances. Free stopped being a duration and became a quantity. The better products spend that quantity deliberately: the free experience shows the intelligence working, once, on the user's own data, well. Enough to prove the value. Not enough to run a business on.
Credits also changed what the pricing page has to explain, which is a product problem rather than a billing one. I worked through that shift in Pricing After Seats.
How should you design free?
Gate by value, not by time, because a 14-day trial of a product someone opens on day 13 proves nothing. Keep the cheap, deterministic features free and habit-forming, and spend the inference budget where it converts: the first session, on the user's own data, where a single well-grounded answer does more than a tour ever did. That first impression carries the weight I described in The AI Trust Gap, which makes it the best possible use of subsidized compute. And enforce the limits in the request path, at run time, because a monthly batch job discovers the overspend after it has happened.
Abuse is a line item now
A free endpoint with real compute behind it attracts scripted usage, resellers, and agents running loops. Every assumption that a signup is 1 person with 1 pair of hands now leaks money, quietly and at machine speed. Rate design, identity checks, and anomaly detection stop being security chores and join the unit-economics conversation. Telling a legitimate agent apart from an abusive one is its own activation problem, and I mapped that funnel in When Your Next User Is an Agent.
A feature with a budget
The free tier still does what it always did: prove value before asking for money. The difference is that the proof now has a unit cost, so free has to be designed like a product, owned like a product, and reviewed like a product. Give it a budget, give it a conversion job, and measure it on both.