AI SAAS COST MODEL AND CONTROL WORKSHEET
Published 2026-10-26 by ShowMeYourSaaS
https://showmeyoursaas.com/articles/ai-saas-cost-controls/

HYPOTHETICAL EXAMPLE: All rates are invented USD values, not current vendor quotes. No measured ShowMeYourSaaS operating costs are included.

Input tokens per attempt: 3000
Output tokens per attempt: 800
Invented input price per million: $2
Invented output price per million: $8
Mean billed attempts per completed job: 1.1
Completed jobs per active user/month: 40
Other variable cost per completed job: $0.002
Fixed monthly infrastructure: $20

Attempt cost = (3000*2 + 800*8)/1000000 = $0.0124
Model cost per active user = 40*1.1*0.0124 = $0.5456
100 active users: 4000 jobs, 4400 attempts, model $54.56, other variable $8, fixed $20, total $82.56.
1000 active users: 40000 jobs, 44000 attempts, model $545.60, other variable $80, fixed $20, total $645.60.
100 heavy active users at 400 jobs each: same $645.60 under these assumptions.
If 20 accounts pay for the 100-user workload: $82.56/20 = $4.128 cost per paying account.
Excluded: payment fees, support, taxes, and other business expenses.

YOUR ASSUMPTIONS
Provider/model/version:
Rate source and checked date:
Definition of completed useful job:
Active users / paying accounts / free users:
Input, output, cache, and other billing units:
Attempts per completed job, including failed work:
Normal and heavy-user job counts:
Variable non-model costs:
Fixed costs and excluded expenses:

CONTROLS TO RECORD
[ ] Server-side authorization and per-account quota
[ ] Bounded input, context, output, tool steps, time, retries, concurrency
[ ] Atomic budget reservation, settlement, and interrupted-job reconciliation
[ ] Global spending control and correct provider/account/workspace scope
[ ] Alerts distinguished from enforced stops
[ ] Request ledger reconciled against provider reports
[ ] Quality and total cost per useful completed job checked together
[ ] Largest input, concurrent quota edge, timeout, retry loop, and exhaustion rehearsed
[ ] User input preserved and recovery explanation tested

Review date and owner:
Observed differences from forecast:
Next adjustment: