In this article

The cheapest model call can be an expensive way to finish a customer’s job. A call may need a retry, a second tool, a human check or a complete restart. The customer still receives one useful result, if they receive one at all.

Price an AI feature using the cost of the accepted task. Keep the call-level costs underneath it, where they explain why the unit changes. This makes a pricing discussion specific: how much work does the system perform, how often does it succeed, and what does the customer actually pay for?

Define the result before counting the cost

For a document-processing feature, “completed” might mean the required fields were extracted, passed validation and were accepted into the destination workflow. Receiving a response from an API is an earlier step. A technically successful response containing unusable fields does not satisfy that customer outcome.

Write the acceptance rule so another person can apply it to a task without seeing the model’s confidence or the customer’s bill. Include partial outputs, abandonment and duplicate submissions. If a task is billed only once, all retries should remain connected to that task’s identifier.

Also decide how a later correction affects the count. A result accepted on Monday and rejected on Tuesday can change both revenue eligibility and support cost. Keep the reporting window explicit rather than quietly moving rejected work out of last week’s denominator.

The billing unit describes what the customer pays for. The cost ledger should explain the work required to deliver that unit.

Build a ledger across the whole attempt

A useful task ledger records the task, its attempts, model usage, external tool charges, review time and final outcome. It should be possible to trace a costly task back to the work that caused the cost.

Anthropic’s pricing documentation separates input, output and cache-related pricing categories. Those are relevant cost dimensions, but the following example uses entirely hypothetical rates. It does not recommend a model or reproduce a provider’s current prices.

Assume 1,000 customer task requests generate 1,300 billable model attempts. Each attempt consumes 4,000 ordinary input tokens and 1,000 output tokens. There are no cache hits in this baseline. Use fictional rates of $2 per million input tokens and $8 per million output tokens, plus a $0.01 external-tool cost per attempt.

A human reviews 100 of the requests for two minutes each at an assumed loaded hourly cost of $30. After all attempts and reviews, 800 tasks meet the acceptance rule. The other 200 do not generate customer revenue.

Cost in the illustrative batchCalculationAmount
Input tokens1,300 × 4,000 ÷ 1,000,000 × $2$10.40
Output tokens1,300 × 1,000 ÷ 1,000,000 × $8$10.40
External tools1,300 × $0.01$13.00
Human review100 × 2 ÷ 60 × $30$100.00
Total included delivery costSum of the rows$133.80

The included cost per accepted task is $133.80 ÷ 800 = $0.16725, or about 16.7 cents. Dividing by 1,300 attempts would answer a different question. Dividing by 1,000 requests would spread cost across work that, under this offer, cannot be billed.

The estimate excludes fixed engineering costs, general overhead, taxes, payment processing, storage and any unlisted tools. Add material costs before treating the result as a complete business margin. The example’s purpose is to keep the cost numerator and the revenue denominator consistent.

See what a lower success rate does

Suppose the customer price is a hypothetical $0.50 per accepted task. The baseline earns $400. Subtracting the included delivery costs leaves $266.20, a contribution margin of 66.55% on that revenue definition.

Now hold the same $133.80 cost constant but suppose only 600 tasks pass. Revenue falls to $300, cost per accepted task rises to about 22.3 cents, and contribution margin falls to 55.4%.

This sensitivity is deliberately simple. In practice, lower success may also cause more retries and review, pushing the cost numerator upward while the accepted-task denominator falls. That combination is why an average cost per model call can be reassuring at exactly the wrong moment.

Track acceptance and cost together by task type. A language, document format or integration that performs poorly can disappear inside an overall average dominated by easier jobs. Segment using categories the team can act on; an exhaustive dashboard of tiny groups is less useful than a few reliable comparisons.

Give the expensive tail its own row

Consider another variation on the baseline. Fifty difficult tasks each need one additional attempt using 20,000 input tokens and 5,000 output tokens at the same hypothetical rates. Each also incurs another one-cent tool charge and four additional minutes of human review. The final accepted count remains 800.

Those extra attempts add $4 in token cost, $0.50 in tool charges and $100 in review. Total batch cost rises to $238.30. At the same $400 revenue, contribution margin drops to approximately 40.4%.

The expensive tail is largely review work in this example. A model-price negotiation would barely address it. Better validation, clearer task boundaries, a customer correction step or a different review policy could be more relevant interventions. Which one works requires evaluation against the acceptance rule, including the quality the customer receives.

Do not exclude the difficult jobs merely to make the unit economics look attractive. Either design a different service for them, price their additional work transparently, or explain which tasks the feature cannot accept.

Estimate first, reconcile after delivery

Preflight token estimates can help identify obviously large inputs. Anthropic describes its token-counting result as an estimate; actual input usage can differ. Its token-counting endpoint also does not simulate prompt caching. Treat estimates as planning inputs and reconcile the batch against actual reported usage and charges.

For caching, keep ordinary input, cache writes and cache reads separate. Do not apply a blended discount to every request merely because caching is enabled. Observe which requests reuse eligible content and whether the reuse survives the timing and shape of your actual workload.

Set an operating budget for attempts and review, with a defined outcome when the budget is reached. That might be a partial result requiring customer action, a delayed specialist review or a clear failure without a charge. The product should explain the outcome instead of silently continuing to spend.

A useful launch decision names the acceptance threshold, included costs, maximum work per task and the customer charging rule. Recalculate with actual tasks after release. The price becomes defensible when the team can show what a completed job requires, including the work hidden behind its successful final response.

Sources

Something we should clarify?

Suggest a correction