A free tier is one of the best growth engines in software — and in AI, one of the easiest ways to lose money without noticing. In classic SaaS a free user costs fractions of a cent, so free is essentially free. In an AI product a free user runs real inference on every click: tokens you pay for, with zero revenue to offset them. The question isn't whether to offer a free tier, it's knowing exactly what it costs.
A free user is 100% cost
For a traditional app, a free user consumes a little storage and bandwidth — near-zero marginal cost, which is why free tiers became a default. AI breaks that. Every free request is input and output tokens billed by your model provider, and none of it is covered by a subscription. Your free tier is a real, metered expense priced in tokens — and it scales with usage, not with signups. A thousand idle free accounts cost nothing; a hundred active ones can cost more than a paid plan.
Estimate the cost of one free user
It's the same math as any customer, minus the revenue: free-user cost = their monthly tokens × your model's rate. The two things that move it are how much a free user does and which model you serve them. The model choice dominates — the gap between a cheap tier and a premium model is more than an order of magnitude for identical usage.
A free user running 200k input + 80k output tokens a month costs about $1.30 on GPT-4o ($2.50 / $10 per Mtok) — but only about $0.08 on GPT-4o mini ($0.15 / $0.60). The default model you route free users to is the single biggest lever on what your free tier costs.
How many free users does one paying customer carry?
That's the ratio that decides whether a free tier is sustainable. If a paying customer nets, say, $37 of gross margin a month and a free user costs $1.30, one paid customer can subsidise about 28 free users before the free tier drags your blended margin negative — or roughly 460 of them if free users run on a cheap model at $0.08 each. Put your real conversion rate next to that ratio: if you convert 1 in 20 free users to paid, a $1.30 free cost needs each paid customer to comfortably cover ~20 free ones. On GPT-4o that's tight; on a cheap model it's easy.
See the cost of each model
Keep the free tier from eating your margin
- Default free users to your cheapest capable model — the biggest lever by far.
- Cap free usage with a hard monthly limit on requests or tokens, not a vague 'fair use'.
- Trim context and output length for the free tier specifically; free users don't need your longest prompts.
- Rate-limit and deduplicate — free tiers attract bots and scripted abuse that quietly run up tokens.
- Track free-tier cost as its own line and alert when it grows faster than conversions to paid.
Free isn't free — but it's knowable
A free tier is a legitimate, powerful acquisition channel when you price it on purpose: know the cost per free user, the number of free users a paying customer can carry, and the conversion rate that makes the trade pay. The failure mode is treating AI free like classic SaaS free — assuming it's ≈$0 when it's actually a token bill that grows with every active signup. Measure it, cap it, and it becomes a growth lever instead of a leak.
Model a free user in the free calculator — set the price to $0 and enter typical usage to see the monthly cost — then connect Stripe and your LLM cost with MarginWard to watch your real free-tier cost against paid margin. Figures above use indicative public rates and illustrative usage.
Go deeper