A model name and token total are not enough to reproduce cost. Grow or Die selects a versioned server-side price using the provider, exact model, event time, service tier, and every non-zero usage meter. If one observed meter has no trusted rate, the call remains unpriced.
The lookup key is more than a model name
Each event retains its provider, exact model ID, operation, service tier, event time, input and output tokens, cache categories, reasoning tokens, and supported tool counts. Aliases are normalized only when the mapping is explicit; a similar-looking model name is not accepted as proof of the same price.
Price the call with the rate effective at call time
Every catalog version has an effective timestamp and provenance ID. A call before a price change keeps the earlier rate. A later catalog update never rewrites historical cost as though the new price had always applied.
Observed call cost = sum of non-overlapping billable quantities × their trusted rates effective when the call occurred.
Separate billable quantities from usage breakdowns
A usage field is not automatically an additional charge. For OpenAI, reasoning tokens are already included in output tokens: do not add them a second time. If output is 1,000 tokens and its reasoning breakdown is 400, bill 1,000 output tokens, not 1,400. Preserve the breakdown for analysis.
Likewise, remove cache-read tokens from inclusive input totals before pricing uncached input. Anthropic cache reads and TTL-specific writes use separate reported categories. Service tier and batch or priority rates select a rate, not an extra copy of the same tokens. Supported tool calls may be separate charges. Normalize according to the provider’s usage contract and retain the catalog version.
One missing rate blocks a complete total
If a call contains a non-zero billable category with no verified rate, calls and tokens remain visible but cost and profit stay unavailable. That is safer than publishing a partial amount that looks complete.
Automatic public-price discovery is provisional evidence. Provider billing remains the reconciliation source for credits, hidden retries, negotiated rates, and unsupported tool charges.
What this evidence can and cannot prove
The catalog calculates cost from usage visible to the customer's server. It does not replace the provider invoice, infer charges for missing usage, or prove that provider-side retries and credits are complete. Reconcile each provider and billing period before treating attributed cost as the full bill.
Primary references
Versioned prices only matter next to real usage.
- AI cost per user — Price the usage with the rate effective at event time.
- OpenAI cost tracking — Unknown models stay unpriced instead of silently costing zero.
- LLM cost per user calculator — Test today's rates without claiming they are a bill.