AI cost per outcome is a feature's AI cost divided by the results it delivers, such as resolved tickets, processed documents or accepted answers, counting every request, retry and fixed cost it took to get them. Cost per request and per task are steps on the way: the request is what the provider bills, the task is what a person sets off, and the outcome is what the business is paid for. The general method for unit costs is in unit economics and unit costs; this page covers what is different for AI.
How it works
Three denominators. Each answers a different question, and only the last can be set against the value a feature creates.
| Denominator | Counts | Source | Answers | Distorted by |
|---|---|---|---|---|
| Request | every model request | the provider's usage data, each response's usage | the effect of model, prompt and price | a varying number of requests per task |
| Task or conversation | one person's action with all its requests | the application's trace | the cost of one use | abandoned and repeated tasks |
| Outcome | a result achieved | the application's business event | comparison with price, value and margin | how the outcome is defined and counted |
The FinOps Foundation traces the same progression for generative AI in the Unit Economics capability of its FinOps Framework (CC BY 4.0): from a cost per token at the start to measures tied to an outcome.
The numerator. The cost of tokens by type (input, cache, output, reasoning) is the core; how it arises is covered in tokens, requests and context. On top of it come:
- Retrieval: embeddings and the vector store behind the documents loaded into the context (retrieval, embeddings and vector databases).
- Tools: server-side tools are charged per use. Anthropic lists web search at USD 10 per 1,000 searches, plus the tokens of the results that enter the context; a search that errors is not billed.
- Retries and failures: a retry that completes is billed again. Whether a failed attempt costs anything depends on the provider: Microsoft states that the HTTP status alone does not decide; Anthropic does not bill batch requests that errored, were cancelled or expired; OpenAI bills the requests that completed within an expired batch.
- Abandoned tasks: requests after which the person gives up cost money without an outcome.
- Evaluation: test runs and grading are model requests too. Amazon Bedrock, for one, bills a model evaluation as inference on the chosen model, and a judge model's tokens (LLM-as-a-judge) at on-demand prices.
- Fixed cost: a fine-tuned deployment in Microsoft Foundry incurs hosting charges while it is deployed, even at low usage, and Provisioned Throughput on Amazon Bedrock is charged per hour per model unit. Per outcome, this cost falls as volume grows.
The calculation:
Cost per outcome = (token cost of all requests + tools + retrieval + evaluation + share of fixed cost) / number of outcomes in the same period
The numerator holds every task of the period, including those without an outcome. That is exactly what separates cost per outcome from cost per successful task.
Counting outcomes. The provider knows nothing of outcomes. The application writes an event when one occurs (ticket closed, document processed, answer accepted), carrying the same task or trace ID as the requests. For each request it logs the model, tokens by type from the response's usage field and the request ID, which Anthropic returns as a header on every response. Through the task ID, request costs sum to tasks; through the event, to outcomes. Which features carry which cost is settled beforehand by AI cost allocation.
Reading it over time. Unit cost moves without any change in volume: a new model with different rates or a different tokenizer, a longer system prompt, a changed price list, a falling cache hit rate. Every response names the model it used; logging model and prompt version as well lets each step in the series be traced to a change.
What it means for your cost
Total cost of a successful AI feature rises because people use it, and that alone says nothing. Cost per outcome says whether each result is getting cheaper or dearer, which makes it the figure a feature owner steers by. Set against the price or revenue per outcome, it becomes part of the technical gross margin; how cloud cost enters the margin is covered in the article on cloud costs in COGS, and how such a figure becomes a measure with a fixed formula in cloud spend KPIs.
The three denominators complement each other. Cost per request shows whether a model or prompt change is cheaper; cost per task shows what one use costs, the basis for usage-based pricing; cost per outcome shows whether the feature pays for itself. For agents, whose number of steps per task varies, the first two often diverge (agents and tool calls).
Because single tasks can cost many times the others, a distribution belongs beside the mean: the median and a high percentile per feature and month show whether a rise comes from every task or from a few.
Where it goes wrong
- Cost per request where requests per outcome vary. Unit cost stays flat while the invoice rises. The cause: an agent needs three requests for a task one time and thirty the next; the denominator grows with the problem and hides it.
- Failed attempts left out. Cost per outcome comes out below the invoice divided by outcomes. The cause: only tasks that produced an outcome were counted, or only the last request in a chain of retries.
- A mean over a long tail. The average looks acceptable while a few tasks consume a large share of the budget. The cause: a distribution with outliers is reduced to one number.
- A denominator nobody counts reliably. The series jumps for no visible reason. The cause: there is no event for "answer accepted", or the definition changes without anyone marking the series.
- A model change read as a volume effect. Cost per outcome rises after a release and the cause is sought in volume. The cause: model and prompt version are missing from the log, so the step cannot be traced to a change.
Sources
Read on 2026-10-16. FinOps Foundation Unit Economics (CC BY 4.0); Anthropic pricing, including web search, batch processing, errors and request IDs and Messages API; OpenAI Batch API and Responses API: create a response; Microsoft plan and manage costs for Microsoft Foundry; AWS Amazon Bedrock pricing.