How to calculate cost per accepted AI video: worked examples

A cheap generation is not necessarily a cheap finished video. The relevant unit might be a five-second product shot, a ten-second face-swap clip or a complete ad. Define that unit first, then keep all attempts and the labor needed to finish it. This guide provides a calculation worksheet and a repeatable comparison method; Magic Hour publishes it and is one possible production workflow.

The formula and the numbers to collect

Total production cost = generation charges + paid retries + editing and review labor + required assets or audio + allocated tool costs. Divide that total by the number of accepted deliverables. Keep cash spending, credit usage and estimated labor in separate columns until you have a documented basis for translating them into the same currency.

For a clip comparison, also calculate cost per accepted second: total production cost divided by the duration of the accepted clips. This helps expose differences in output length, but it does not make all seconds equally valuable. A usable two-second product reveal and a ten-second establishing shot are different tasks and should remain separate cohorts.

Record attempts, successfully rendered outputs, accepted outputs, accepted seconds, billed credits or dollars, refunds, operator minutes and final resolution. Also record the source, prompt, model or tool version, generation settings and test date. A file that rendered successfully but fails the brief is a rejected output, not an accepted one.

1. Define a deliverable before generating

Write a pass/fail brief that another reviewer could apply. For a product shot, require the correct package, readable approved text, one intended movement, the destination aspect ratio and a usable export. For dialogue, require the approved words, comprehensible audio and acceptable lip timing. For face swap, require the intended identity through the whole selected clip.

Decide whether you are buying a raw clip or a finished deliverable. If the production workflow always adds captions, music and a logo in an editor, include that finishing work in every option. Do not penalize one tool for missing captions while approving a raw result from another. Do not count several variants of one accepted clip as independent successes unless the brief genuinely requires separate deliverables.

Use the 42-point AI video QA checklist to make the acceptance criteria concrete. Freeze those criteria for a test round. If the client changes the required package or dialogue, start a new round instead of silently changing the denominator.

2. Count failures, retries and refunds correctly

Every charged attempt belongs in spending, including a render that completes with an unusable label, an identity change or poor motion. A processing error may have a different billing treatment from a completed result you reject for quality. Record the actual account charge and confirmed refund; do not assume the two cases are identical.

Keep request count separate from generated outputs. A timed-out client request can correspond to a job that was accepted and is still processing. Before resubmitting, find whether the original project exists. Duplicate submissions can inflate costs and make a supposedly fair benchmark measure your integration error instead of the underlying tool.

For API work, the Magic Hour API billing overview is the billing reference, and the media API evaluation guide is a starting point for defining the test. Inspect the chosen endpoint and current quote. An API request, a successful render and an accepted creative are three different counts.

3. Translate credits without inventing a dollar rate

Credits are provider-specific units. One credit from one service is not interchangeable with a credit from another. For each workflow record the displayed cost and its settings, then preserve the pack or subscription terms that funded those credits. Different models, resolutions, durations or enhancements can have different rates.

A simple purchased pack can provide an effective unit rate: pack price divided by credits received, with taxes and currency handled consistently. A subscription needs a declared allocation method. Dividing the monthly fee by all included credits assumes a usage pattern; it is not automatically the cash cost of a particular clip. Report the method beside the number.

Unused balance is not the same as money refunded, and buying a pack can create a larger upfront commitment than the credits spent on one test. Show both the cash required to start and the allocated cost of the experiment. If an existing subscription would be paid anyway, distinguish incremental generation cost from fully allocated production cost.

4. Include the time needed to finish the video

Time prompting, input preparation, reviewing rejected takes, fixing a result, adding required audio or captions, exporting and approving. If one model needs twenty minutes of cleanup while another passes after two minutes, omitting labor can reverse the apparent winner. Use the same hourly rate and timing convention for all compared routes.

Separate active operator time from unattended waiting. Queue and render delays affect turnaround; they do not always consume paid editing time. Record both. If you are waiting at the screen because the deadline prevents other work, state how that waiting time was allocated rather than quietly folding every elapsed minute into labor.

A reusable setup cost can be shown separately or allocated over a declared production volume. For example, a template built for fifty expected ads should not be hidden as free work, and its entire cost should not be assigned to one ad without explanation. State the volume assumption and recompute if production stops early.

A break-even check for a more expensive workflow

The most you can justify paying to save operator time depends on the time actually saved. In an illustrative case, twelve minutes saved at $30/hour are worth $6. If a different generation workflow costs an additional $2 per accepted deliverable and saves those twelve minutes, its estimated net saving is $4. If it saves only two minutes, the time value is $1 and the additional $2 costs more than it saves.

Apply this calculation to accepted deliverables rather than individual attempts. Measure the preparation and review burden too. A workflow that renders quickly but creates more rejected outputs may save no operator time. Use your own recorded time and labor rate, and show the assumption when the estimate is uncertain.

Worked example: cheaper attempts can cost more to finish

The following numbers are fictional and demonstrate the formula. Both routes aim to deliver two approved five-second product clips. They use the same brief and labor rate; these are not measured results for Magic Hour or any competitor.

Route A: eight attempts at $1 each cost $8. Twenty-four minutes of finishing at $30/hour cost $12. Two accepted clips give ($8 + $12) ÷ 2 = $10 per accepted clip, or $2 per accepted second.

Route B: four attempts at $2 each also cost $8. Eight minutes of finishing at $30/hour cost $4. Two accepted clips give ($8 + $4) ÷ 2 = $6 per accepted clip, or $1.20 per accepted second. The higher attempt price wins in this illustration because it requires less finishing, not because price itself proves quality.

If Route B required a new $20 subscription solely for this test, include that commitment. Charging all $20 to these two clips gives ($8 + $4 + $20) ÷ 2 = $16. If you allocate the subscription across more production, disclose the denominator. This is why a comparison should show both incremental cost and the cost of entering the workflow.

Worked example: zero usable output is not a bargain

Suppose five attempts cost $5 but none preserve the approved package label. Accepted outputs equal zero. The result is “$5 spent, zero accepted deliverables,” not $0 per usable video and not $1 per usable video. Stop, change the source or simplify the shot before spending on unchanged retries.

If the sixth attempt passes after a $1 charge and ten minutes of review at $30/hour, the observed cost is ($6 + $5) ÷ 1 = $11 per accepted clip. Keep the five rejected attempts in the record. Reporting only the successful attempt would erase the work required to get there.

A worksheet you can copy

Job definition: deliverable, duration, resolution, aspect ratio, required content, source permissions and acceptance criteria. Configuration: tool, model/version, plan, date, prompt, source assets and settings.

Observed counts: requests submitted, jobs accepted, processing failures, generated outputs, quality rejections, accepted deliverables and accepted seconds. Observed spend: credits charged, credits refunded, cash payments, asset costs and allocated subscription or setup cost.

Time: input preparation, prompting, active review, correction, export, approval and unattended wait. Result: acceptance rate, incremental cost per accepted clip, fully allocated cost per accepted clip and turnaround. Mark missing fields as unknown rather than filling them with estimates presented as measurements.

How to run a fair tool comparison

Choose several representative briefs from the work you actually do. Include difficult cases such as a product turn, a profile face, a hand crossing the subject or a spoken proper name when those occur in production. Use the same authorized source assets, duration, target quality and acceptance rules wherever the tools support them.

Use the settings available to a normal customer and record them. A special beta, enterprise configuration or vendor-curated output should not be mixed into a standard self-service cohort. If two tools cannot produce the same required format, record the incompatibility rather than quietly lowering the brief for one of them.

Retain all attempts and have a reviewer apply the criteria consistently. If possible, hide the provider label during creative review. Record why each failure occurred. Summarize each task separately before calculating an overall result; a provider that handles talking heads well may still fail product labels.

There is no universal sample size that proves a winner across every workflow. Publish the number of briefs, attempts and accepted outputs, the collection period, exclusions and limitations. A tiny pilot can identify a serious failure or a useful candidate, but it cannot establish a market-wide quality ranking.

When to regenerate, edit or change workflow

Regenerate when a specific input or setting can address the defect. Edit when the accepted motion is usable and the remaining change is bounded, such as adding verified text in a conventional editor. Switch workflows when the model repeatedly breaks a hard requirement. Keep repair time in cost whichever path you choose.

Do not buy additional attempts before diagnosing a repeated failure. Use the prompt troubleshooting guide to isolate source, prompt, settings and capability. Exact typography, product facts, permissions and approved dialogue remain hard gates even if a visually impressive clip is cheap.

Official references and limits

Check Magic Hour’s current model reference and the relevant endpoint for available settings; check the billing overview for billing behavior. To create the illustrative product-shot workflow, use Image-to-Video. These links establish product references, not the fictional outcomes used in the arithmetic examples above.

This article contains no measured cross-provider benchmark. Its calculation method is reusable, but any published provider comparison needs retained aggregate data, dated sources, clear denominators and reproducible settings. For choosing a tool today, combine this worksheet with the AI video generators guide.