The free tier as customer acquisition cost
Updated September 1, 2026 · first published September 1, 2026
Inference is not free to serve. When a provider hands you a free tier, generous rate limits during a preview, or an unbranded endpoint that costs nothing to call, that compute is paid for — by them, out of a budget that has a name. It is customer acquisition cost, and it behaves like every other CAC line in every other business.
This is the most useful lens a finance team can apply to LLM pricing, because CAC has predictable dynamics. It is spent to win a cohort. It has a payback period. And it stops when the cohort is won.
What a free tier is buying
Three things, usually at once. Evaluation: engineers try the model without a procurement conversation, which is the only way to get into a stack that already has an incumbent. Data on real workloads: what people actually ask, at what shapes and lengths, which is worth more than the compute costs. A price anchor: if your engineers get used to a capability at zero, every competitor's paid rate for that capability now reads as expensive.
None of those goals survive the cohort being won. That is the part teams misprice.
The signal to read
The generosity of a free tier tells you roughly where that provider thinks it is in the acquisition cycle. Aggressive free limits on a new model mean the lab is buying distribution and has not yet decided what the thing is worth. Free limits quietly tightening means the acquisition phase is closing and the rate card is next.
Watch the shape of the tightening, not just the headline. Rate limits fall before prices rise, and context limits, concurrency, and which models are included in the free tier all move before anyone publishes a new price.
How to actually use it
Never let a free tier into your unit economics. If a workload only clears its margin because part of the inference is free, you do not have a margin, you have a subsidy with an expiry date you do not control.
Price free usage at a shadow rate. Pick the provider's nearest paid tier, apply it to your free traffic in telemetry, and carry that number in the forecast. When the free window closes you already know the size of the hit, on the day it happens rather than at the end of the month.
Measure switching cost while it is still cheap to leave. The purpose of a free tier is to make leaving expensive. Prompts get tuned to one model's quirks, evaluations get built against its outputs, tool schemas get shaped to its behaviour. Keep a second provider wired into at least one route so the answer to a price change is a config change and not a project.
Related
Related
Want this applied to your own LLM spend? FinOps LLM runs a free audit of your AI costs and shows where the savings are. Book free audit →