The decision
A provider’s per-token price is one input to unit economics. Your product’s cost depends on input length, output length, retries, failed generations, caching, storage, payment fees, and support. Estimate from observed workflow runs and state which values remain assumptions.
A worked example
Suppose an illustrative feature costs $0.03 for one attempt, succeeds usefully on 80% of attempts, and requires two attempts for some users. Dividing cost by successful results gives a more relevant figure than quoting the cheapest call. For a one-time digital kit, model cost may be zero but fees, refunds, email, hosting, and support still affect contribution.
How to put it into practice
- Record cost for every attempt, including failures and retries.
- Calculate cost per useful result and per active customer, not just per request.
- Use actual processor fees and refund behavior from your account instead of assuming a universal rate.
- Separate variable costs from fixed monthly costs and label both revenue and contribution clearly.
A failure to plan for
Gross revenue is not profit. A $1,000 revenue day can have very different economics depending on product delivery cost, tax treatment, acquisition cost, and refunds. Keep the target and the margin model separate.
Try it on your project
Collect twenty representative runs and enter observed costs into the tool. Compare a typical case with a long-input case and a retry-heavy case. Set a spending limit that remains safe in the worst case you have measured.
Keep the next step small
Use the free demand scorecard or planning tools to make your assumptions explicit. The $19 launch kit brings the blueprint and seven editable worksheets together.