Skip to content
sinc-LLMBook a 30-minute call

AI Cost Per Query: A Worked Worksheet for Tokens, Retries and Human Review

Calculate AI cost per query with a worked worksheet for tokens, repeat calls, human review and accepted results.

On this page

By Mario Alexandre · Updated October 16, 2026 · 7 min read

Your AI feature can look cheap in the API dashboard and expensive in the review queue. In the fictional workload below, 1,000 requests use $6.72 in model calls, then reach $166.72 after human review and allocated infrastructure. Those illustrative costs exclude development, support, taxes and other unlisted costs.

Before choosing a cheaper model, calculate three separate figures: model cost per request, operating cost per request and cost per accepted result. That shows which part of the workflow deserves attention and prevents extra calls from being counted as new requests.

Use the worked worksheet to replace the hypothetical prices, token counts, retries and review time with your own figures. You will leave with a cost calculation you can explain and an assumption to test next.

Decide what one query means

A user asks one question. Your application retrieves documents, calls a model, rejects an incomplete answer, and calls the model again. That is one user request and two model calls.

If you divide the bill by model calls, the result describes a call. It does not describe the cost of serving that user request.

Use a stable request ID to connect the calls, retrieval work, retries and review that belong together. Choose a fixed reporting period and one currency. Record whether your denominator includes failed requests. For an operating-cost calculation, it should: unsuccessful work still consumes resources.

Then define acceptance. An answer could be accepted automatically under a documented rule, approved by a reviewer, or confirmed to resolve a task. Those are different standards. Name the one you use before comparing costs.

Calculate the model charge for each call

For a simple text call without caching or other fees:

call cost = input tokens × input price per million / 1,000,000
          + output tokens × output price per million / 1,000,000

Real billing can have more categories. For example, Anthropic lists separate input, output and prompt-cache charges, with additional conditions for some services. Use the billing categories and rates that apply to the exact model and request mode you ran. Claude pricing documentation.

Do not count the same tokens twice. If a usage field includes cached input, first understand how that provider separates billable categories. Sum mutually exclusive quantities at their applicable rates, rather than adding a cache charge on top of an already inclusive input charge.

Record the model identifier, rate source, currency and date alongside the calculation. If requests use several models, calculate each call at its own rate before summing. A single blended rate can hide an expensive fallback path.

Estimate the token-charge component with the LLM cost calculator, then add repeat calls, review and infrastructure in this worksheet.

Work through a 1,000-request example

The following values are illustrative assumptions. They are not current vendor prices, a customer result, or a claim about a typical application.

Assume every model call uses 1,200 input tokens and 400 output tokens. Set hypothetical prices of $2 per million input tokens and $8 per million output tokens.

Component Calculation Cost
Input for one call 1,200 × $2 / 1,000,000 $0.0024
Output for one call 400 × $8 / 1,000,000 $0.0032
One call $0.0024 + $0.0032 $0.0056
Initial calls 1,000 × $0.0056 $5.60
Additional calls 200 × $0.0056 $1.12
All model calls 1,200 × $0.0056 $6.72

The additional calls belong to the original 1,000 requests. They do not create 200 new user requests. In this example they use the same token counts as the initial calls; replace that assumption with actual usage when available.

Now include human review. Suppose 300 requests each need 45 seconds of review, and the fully loaded review rate is an assumed $40 per hour:

review hours = 300 × 45 / 3,600 = 3.75 hours
review cost  = 3.75 × $40 = $150

Allocate another hypothetical $10 to retrieval and serving infrastructure for this workload. Keep the allocation method visible; a shared monthly bill is not automatically attributable to one feature.

Cost layer Total
Model calls, including additional attempts $6.72
Human review $150.00
Allocated retrieval and serving $10.00
Defined operating cost $166.72

The model cost per user request is $6.72 / 1,000 = $0.00672. The defined operating cost per request is $166.72 / 1,000 = $0.16672, or about 16.7 cents.

This calculation excludes engineering development, support, taxes and any other cost not explicitly listed. Add those when they belong to the decision you are making. Do not quietly call a partial operating estimate the total cost of ownership.

Check the cost of a useful result

Suppose 800 of those requests meet your acceptance rule. The remaining 200 still belong in the cost numerator:

cost per accepted result = $166.72 / 800 = $0.2084

That is about 20.8 cents per accepted result. It is not the same as dividing by all requests, and acceptance does not necessarily mean the underlying business task was resolved.

For a deeper treatment of the denominator, continue with AI cost per resolved task. Use that distinction when a cheaper answer might create another support ticket or another review cycle.

Find the assumption that changes your decision

In the fictional example, review is the largest listed cost. Switching to a cheaper model would address only part of the $6.72 model charge.

Consider a sensitivity check instead. If the same 300 reviews took 15 seconds each, review would cost $50. Keeping every other assumption unchanged would produce an operating total of $66.72, or $0.06672 per request.

The arithmetic shows a $100 difference. It does not show that review time can safely be cut. To test that possibility, investigate why reviewers spend time: missing citations, inconsistent formatting, ambiguous acceptance criteria, or genuinely difficult decisions. Improve the cause you observe, then check both review time and accepted-result quality.

The same discipline applies to retries. Separate transport failures, validation failures and answers rejected for quality. Each may need a different fix. A retry count by itself does not identify the cause.

Build a ledger you can audit

Keep two linked records. A call record contains request ID, attempt ID, model, billable token categories, prices and any separately billed services. A request record contains acceptance status, review time and allocated operating costs.

This separation prevents review time from being counted once for every retry. It also lets you inspect expensive requests rather than relying only on an average.

Before trusting the result:

  1. Reconcile included call charges against the provider’s usage or billing records for the same period.
  2. Investigate missing request IDs and calls that do not join to a request.
  3. Check whether failures, fallback models and repeated attempts are included.
  4. Explain cost allocations and exclusions.
  5. Compare the result across similar tasks and acceptance rules.

A useful dashboard can show the average alongside the distribution and the expensive outliers. Keep latency and quality beside cost so an apparent saving cannot hide a worse user experience.

Make one measurable improvement

Choose one workflow and calculate a baseline using the worksheet. Replace illustrative prices and token counts with retained usage, and measure review time rather than guessing when possible.

Then change one identifiable part of the workflow. Compare cost, acceptance, review effort and response time on comparable cases. Retain the failures as well as the successes.

If you need help tracing costs across model calls, integrations and human review, review the AI Systems Review engagement. A redacted request trace and a clear acceptance rule make a useful starting point.

Need help applying the ledger to a production workflow? Review the AI Token Cost Engineering scope.