Use active users and real feature mix
A monthly API invoice divided by signups hides the spending pattern. Start with monthly active users, then count calls per active user for each AI feature. A summarization call, chatbot turn and document analysis request have different token mixes and prices.
Price one feature at a time
For each feature, record the exact model, average input tokens, average output tokens and per-million rates. Cost per request is (input × input rate + output × output rate) / 1,000,000. Multiply by calls per active user. For example, 20 calls at $0.01 each cost $0.20 per active user before other features.
Include tails and failures
Averages alone can hide power users and long documents. Measure p50 and p95 token use, retry rate and failed calls. Put caps on maximum output and context where product quality allows. If usage is concentrated among a small group, model that group separately so a generous free tier does not surprise you.
Connect cost to margin
Suppose a paid plan earns $10 monthly per active user and AI usage costs $0.80. AI alone uses 8% of revenue; payment, hosting, support and acquisition costs still need room. Monitor cost per successful feature outcome as well as per user. Estimate monthly API spend under several adoption scenarios.