Cost per action (CPA) is the number that decides most AI economics. Not cost per token or cost per API call — cost per unit of value delivered to the user. When CPA fits within the value delivered, the feature is viable. When it does not, no amount of prompt engineering will save it.
- 01Compute CPA for a defined feature.
- 02Distinguish CPA from cost per token.
- 03State the pricing implications.
The calculation
CPA = (average tokens per action × per-token cost) + amortized infrastructure per action + non-AI operational cost per action. Not: token cost alone. A feature that looks cheap at the token level can be expensive at the CPA level once retries, chained calls, and downstream storage are counted.
The implications
Price the feature above CPA with margin. Free-tier or freemium features must have low enough CPA to survive abuse at scale. High-CPA features (deep reasoning, long-context, multi-agent) require pricing that reflects the cost — usually a premium tier or usage-based billing.
A worked CPA
Feature: summarize a 20-page document. Input: ~20K tokens. Output: ~500 tokens. On GPT-4o: 20K × $2.50/M + 500 × $10/M = $0.055. Add reranking retries (~10% surcharge) and infra (~$0.01), CPA ≈ $0.07. Free-tier without limits is unviable; a 100/day cap or a per-summary charge is.
- CPA is the economic unit that matters.
- Token cost is a proxy; CPA is the number to price against.