A cost model for AI features
A spreadsheet that turns tokens and volumes into a monthly number.
What we found
Most surprises come from retries, long contexts and staff usage growth — not from the headline price per token.
Teams usually estimate the cost of an AI feature from the per-token price, then get a bill several times larger. The gap is in the parts nobody models: retries, unexpectedly long inputs, evaluation runs, and the fact that usage rises once the feature is good.
We built a small model that takes volume, average input and output length, retry rate and a growth assumption, and returns a monthly cost with a sensible upper bound.
It is deliberately boring, and we hand it to clients during scoping so the business case is settled before anything is built.
[Publish the sheet here, or link to it — this entry should end with something the reader can actually use.]
Draft — re-run and replace [bracketed] figures before publishing (lib/lab.ts)