How to calculate cost per usable generation

The price of one generation is not the cost of one usable asset. Production cost includes unsuccessful attempts, reruns, correction work, and outputs that look appealing but fail the brief. A meaningful calculation uses accepted deliverables as the denominator. It also states what “usable” means before the results are seen.

This guide provides a measurement method, not a current price table. Prices and credit rates can change, account conditions may differ, and a model route may be revised after publication. Verify current pricing in the product before using any estimate for a budget. Use the comparisons hub to keep candidate inputs consistent and the workflow directory to preserve the production record.

Assumptions

The following example is illustrative and dated 2026-09-04. Assume a team runs a fixed brief for a defined number of attempts, records the value consumed by each attempt under the terms visible to that account, and reviews every output against one acceptance checklist. Assume correction time is included only when the team can estimate it consistently. Exclude unrelated overhead unless it changes between candidates. Treat refunds, failed technical jobs, subscriptions, and bundled credits according to the actual terms in force, documenting the choice rather than silently assigning them a value of zero.

The unit of analysis must remain stable. One accepted storyboard frame, one approved product image, and one finished video shot are different units and should not share a denominator. If an output is useful only as an intermediate source, count it in the workflow that eventually produces the deliverable. Record currency, tax treatment, and whether the estimate uses purchased credits, allocated subscription value, or another internal accounting convention.

Decision criteria

Define an accepted output with observable conditions: correct subject, required composition, adequate integrity, delivery format, and correction effort below a stated threshold. Then track direct generation value, acceptance rate, human review time, correction time, and time-to-delivery. Cost per usable generation equals total relevant cost across the test divided by the number of accepted outputs. When no output passes, report the result as no usable output rather than forcing a misleading finite number.

Compare candidates on the same unit and attempt budget. A model may consume more value per run but achieve a higher acceptance rate. Another may be inexpensive per attempt yet require enough retries and editing to cost more per delivery. Separate cash cost from labor and elapsed time so a team can apply its own rates. The calculation should reveal the tradeoff, not hide assumptions inside one authoritative-looking figure.

Inputs

Collect the exact brief, model or workflow route, settings, output count per run, number of attempts, accepted-output count, and value consumed under the current account terms. Record review minutes and correction minutes separately. Assign each rejected output one primary reason so the team can distinguish unavoidable variance from a correctable prompt or source problem. Save the date because current pricing and credit terms are volatile inputs.

Use a simple table with one row per attempt. Include run identifier, direct value, technical completion status, acceptance result, rejection reason, review time, correction time, and final disposition. Preserve the underlying outputs so another reviewer can audit the acceptance decision. If a subscription covers several activities, state the allocation method. Do not divide the entire fee by one test unless that reflects how the budget is actually managed.

Failure modes

The most serious error is dividing total spend by all generated files. That calculates cost per output, not cost per usable output. Another error is excluding failed attempts or counting a near miss as accepted because it was expensive. Apply the checklist consistently. Small tests can also exaggerate the result: one additional acceptance in a batch of three changes the ratio dramatically. Report the sample size beside every figure.

Labor estimates can create false precision. Reviewers work at different speeds, and correction complexity is not linear. Keep direct value and labor minutes visible as separate measures before combining them. Avoid comparing a highly refined prompt on one model with a first draft on another. Finally, do not copy an old unit price into a permanent calculator. Verify the latest pricing or credit information whenever the budget, account, route, or review date changes.

Limitations

This method describes a dated internal estimate, not financial advice or a universal market price. It cannot predict future service terms, queue behavior, model revisions, taxes, exchange rates, or regional access. Acceptance also depends on the brief and reviewer. Results from simple landscapes do not automatically forecast a product series, complex character motion, typography, or regulated creative work.

Some outputs create value without becoming final deliverables. Exploratory images may clarify art direction, and rejected clips may reveal a useful camera choice. A strict usable-output calculation does not capture that learning value. Record it separately rather than changing the acceptance definition after the test. Likewise, legal clearance, brand approval, and downstream editing may be essential costs even though they are not generated by the model itself.

Next actions

Choose one deliverable type and write its acceptance checklist. Set an equal attempt budget for each candidate from the models directory. Capture the current account terms and date before starting. Run the fixed brief, preserve every output, and record direct value, review minutes, correction minutes, and disposition. Calculate acceptance rate first, then divide total relevant cost by accepted outputs.

Present the result as a range when labor allocation or bundled credits are uncertain. Include sample size, currency, date, and exclusions beside the number. Identify the largest rejection category and test one controlled change that could reduce it. Recalculate only after preserving the original baseline. Set a refresh trigger for any change in pricing, credit terms, model route, account plan, or production brief so the estimate remains a decision tool rather than an outdated claim.