Tutorial · Evidence checked 2026-08-30
How to Calculate AI Video Cost per Approved Finished Second
A practical, evidence-led decision guide. Product capabilities and limits are separated from anything that would require hands-on testing.
AI video pricing becomes misleading as soon as “one generation” is treated as one deliverable. A five-second result can consume different credits by model, resolution, audio mode or platform. It may also be rejected, regenerated, extended, edited and reformatted before one second reaches a customer.
The useful unit is cost per approved finished second. It connects the vendor meter to the asset the business can actually publish.
Build the denominator from the campaign brief
Record the required finished seconds by format, not just the number of videos. A launch might need:
- one 20-second hero film in 16:9;
- three six-second cut-downs in 9:16;
- six four-second product loops in 1:1;
- two language or message variants.
That is 80 required finished seconds before alternate takes. Keep a separate count for each format if cropping or regeneration is required. Do not assume one output can be reframed without visual loss.
Translate the provider meter without pretending credits are comparable
For each exact platform/model/mode, capture:
| Input | Example of the evidence required |
|---|---|
| Price paid | invoice or current checkout, including billing period |
| Included allowance | credits, compute units or seconds |
| Model consumption | credits per second or per generation at selected settings |
| Duration/resolution | the exact deliverable the rate covers |
| Expiry/rollover | official plan rule |
| Failed-job treatment | whether credits are restored and under what condition |
| Web/API separation | separate balance, rate card or contract |
Credits are internal meters. Sixty Kling credits, sixty Krea compute units and sixty Freepik credits have no shared economic meaning. Convert each to currency only after the plan price and eligible allowance are known.
Calculate submitted generation cost
When a provider charges per second:
`submitted generation cost = total seconds requested × price per generated second`
When it charges credits:
`credit value = eligible plan price / usable plan credits`
`submitted generation cost = credits consumed × credit value`
The word “eligible” matters. If a Freepik plan bundles stock assets the team already needs, finance may allocate only part of the subscription to video. If a Runway allowance is shared across a workspace, use the actual credits consumed by the project, not the whole plan or a fictional equal split.
For annual pools, match the period. Dividing an annual allowance by 12 is useful for steady-state planning but can conceal a seasonal campaign's ability to consume a large share at once.
Add rejection and revision
Define:
`technical pass rate = outputs without corruption or platform failure / outputs completed`
`creative acceptance rate = outputs approved / outputs reviewed`
`submitted seconds per approved second = submitted generated seconds / approved finished seconds`
If 600 generated seconds produce 120 approved seconds, every approved second carries five generated seconds of cost. A provider at $0.10 per generated second contributes $0.50 per approved second before editing.
Separate platform failures from creative rejection. A malformed file and a legally unusable likeness are different problems even when neither ships.
Add the labour that cheap generation creates
Track prompt preparation, queue monitoring, review, rights checks, editing, sound, typography, export and project management.
`labour cost = hours by role × loaded hourly rate`
`total approved-second cost = (generation + labour + storage + required seats + rights/licensing) / approved finished seconds`
Example:
- plan allocation: $120;
- overage: $80;
- 8 creative hours at $55: $440;
- 2 legal/brand review hours at $90: $180;
- 160 approved seconds.
Total cost is $820, or $5.13 per approved second. Reporting only the $200 generation spend would hide 76% of the cost.
Compare a creative UI and an API separately
An API adds engineering, storage, retries, moderation handling and observability. A creative UI adds seat, review and handoff costs. Do not transfer the web plan's credit value into the API row unless the official documentation says the balance is shared.
For automated systems, add:
`API operating cost = generation + orchestration compute + storage/egress + human exception review`
Measure retry amplification: submitted jobs divided by unique approved assets. A silent retry loop can dominate spend.
Run a 20-output calibration before forecasting a year
Use one representative brief and 20 attempts per selected model. Preserve prompt, input rights, settings, credits, duration, result, rejection reason and review minutes. Calculate the observed acceptance rate and use a conservative confidence range rather than treating 20 outputs as certainty.
Repeat the calibration for the hardest content class: people, product geometry, text, continuity or camera motion. An average across easy abstract clips will understate campaign cost.
The calculator output procurement needs
The final sheet should show:
- currency per submitted generated second;
- submitted seconds per approved second;
- labour per approved second;
- total cost per approved second;
- monthly capacity at the selected plan;
- the first limit reached: credits, seats, queue, rights or reviewer time;
- unknowns that require checkout, contract or hands-on evidence.
That turns a model demo into an auditable budget.
Primary next step: Compare AI video tools by their real billing boundary and calculate one campaign before choosing a plan.