Token cost is the provider-priced cost of model usage. It is not profit because profit also needs revenue, refunds, processor fees, and a defined unit. To judge a customer or feature, join usage to payments from the same identity and period, and keep unmatched amounts unallocated instead of spreading them across everyone.
A token record answers a cost question, not a business question
OpenAI and Anthropic return usage fields that separate input, output, and caching activity. Those fields make provider pricing possible. They do not contain your customer’s payment, refund, processor fee, acquisition source, or internal account identity.
Even a perfectly priced model call can answer only what this call cost under the selected price schedule. It cannot answer whether the user paid enough, whether the feature caused a purchase, or whether the company made a profit.
The smallest honest profit formula
AI contribution profit = observed payment revenue − refunds − processor fees − attributable AI cost
Apply every term to the same currency, complete date range, and population. If the payment is linked to account A but the model call has no account ID, do not allocate that cost to A based on an average. Show it as unallocated cost until the identity link is observed.
This is an operational contribution metric, not company net income. Payroll, non-model infrastructure, support, taxes, chargebacks, and other expenses remain outside it unless they are explicitly added and labeled.
Worked example: the same token bill, two different outcomes
The figures below are hypothetical. Both accounts generate the same provider-priced AI cost in one complete month, but their payment evidence is different.
| Observed item | Account Cedar | Account Maple |
|---|---|---|
| Payment revenue | $180 | $70 |
| Refunds | $0 | $10 |
| Processor fees | $6 | $3 |
| Provider-priced AI cost | $45 | $45 |
| AI contribution profit | $129 | $12 |
Cedar: $180 − $0 − $6 − $45 = $129
Maple: $70 − $10 − $3 − $45 = $12
The $45 AI cost is real evidence in both rows, but it does not have a profit meaning until it meets the account’s payment facts.
Keep the data boundaries visible
- Token count is not currency. Separate model, input, output, cached-read, and cached-write usage before applying the price effective for that event.
- SDK-priced cost is provisional. Missing calls, provider adjustments, tier rules, or an unknown model can make it differ from the organization invoice.
- Identity coverage controls customer claims. Site-wide cost can be complete while customer-level cost is unknown because account IDs are missing.
- Payment coverage controls revenue claims. A billing total without a customer link cannot establish which user was profitable.
- Observed association is not causation. A costly feature used before payment did not necessarily cause the sale.
Which number supports which decision?
| Evidence available | Safe conclusion | Do not conclude |
|---|---|---|
| Usage only | Which model or feature consumes more tokens | Which customer is profitable |
| Priced usage only | Observed AI cost by model or feature | Business margin or willingness to pay |
| Payments plus site-wide cost | Site-wide AI contribution metric | Customer-level profit |
| Exact account join | Customer contribution for the selected period | Lifetime value or causal impact |
Continue with the full AI unit economics formula , compare gross margin and AI contribution margin , or learn how to find loss-making customers without guessing .