Two numbers that are not the same
Vendors price per second of generated video. You buy finished video. Between them sits everything you discarded.
| Measure | What it counts | Who quotes it |
|---|---|---|
| Price per second | One second of raw output | Every vendor |
| Cost per generated minute | Sixty seconds of raw output | Easy to compute, still misleading |
| Cost per usable minute | Sixty seconds you would ship | Almost nobody quotes it |
| Cost per finished minute | Usable minute plus edit and audio time | Your actual budget line |
The only number that matters: cost per finished minute. Everything else is an input to calculating it, not a price you can plan with.
Regeneration waste is the hidden multiplier
You never keep everything you generate. The proportion you keep — your approval rate — divides directly into your effective cost.
- At a 50 percent approval rate, your real cost is double the headline.
- At a 25 percent approval rate, it is four times.
- Approval rate varies enormously by content type: simple B-roll can exceed 70 percent, specific character work on demanding shots can fall below 20 percent.
This is why a nominally cheaper model frequently costs more. A model at half the price with half the hit rate is a worse deal, and the price comparison will never reveal it.
Worked examples
Illustrative arithmetic using a one-cent-per-second reference rate. Substitute your own rates.
| Scenario | Generated | Approval rate | Effective cost |
|---|---|---|---|
| Simple B-roll, generic prompts | 1 min at $0.60 | 70% | About $0.86 per finished minute |
| Product clips with specific framing | 1 min at $0.60 | 40% | About $1.50 per finished minute |
| Character work across multiple shots | 1 min at $0.60 | 20% | About $3.00 per finished minute |
| Demanding cinematic shots with retries | 1 min at $0.60 | 10% | About $6.00 per finished minute |
Then add editing, which for most teams is the larger number. Generation is increasingly the smaller half of the budget.
What actually drives your cost
- Approval rate — the single biggest lever, and it is a skill variable, not a pricing one.
- Resolution — high-resolution generation costs disproportionately more and is often unnecessary if you refine later.
- Clip length — long clips cost more and fail more often, so retries compound.
- Model choice — less important than approval rate once you are comparing reasonable options.
- Human time — usually the dominant cost at professional quality levels, and the one most often omitted from AI cost comparisons.
Comparing vendors honestly
-
Fix one test set
Twenty prompts representing your real work. Never change them.
-
Run every vendor on the same set
Identical settings, identical resolution targets.
-
Count publishable results
Not pretty results — clips you would actually ship.
-
Divide price by publish rate
That is your comparable number.
-
Add your edit time
If one vendor needs more cleanup, price that in.
Promotional pricing breaks comparisons: launch rates are subsidies designed to win evaluation. Compare on list price or you will be re-doing this analysis when the discount ends.
How to reduce cost without losing quality
- Improve prompts before switching vendors. Better prompts raise approval rate, which beats any price difference.
- Generate shorter clips and cut them together. Short clips fail less and waste less per failure.
- Generate at moderate resolution and refine, rather than paying for native high resolution on every attempt.
- Approve before upscaling. Never spend refinement compute on clips you will discard.
- Reuse what works. Build a library of prompts and reference images with proven hit rates.
A simple cost model you can reuse
Four inputs, one output.
- P = price per second of output at list rate.
- A = approval rate as a decimal.
- E = your edit hours per finished minute, at your hourly rate.
- Cost per finished minute = (P × 60) ÷ A + E.
Run it for each content type separately. A blended average across very different work hides the cases that are actually losing money.