AI video budgeting: measure cost per approved second
API price per generated second hides retries and review. Use separate generation and approval ledgers to learn what each delivered second really costs.
The useful unit is an approved second
AI-video pricing pages usually quote the cost of a generated second. A product team, however, does not ship every generated second. It ships the small fraction that survives creative review, brand checks, editing, and approval.
That difference changes the question from “Which model has the lowest rate?” to “What does one approved second cost after retries and review?” The first number belongs to an API price sheet. The second describes your production system.
What the source articles actually establish
Verified source facts: Johnny Mai’s Alibaba Cloud Community article, published October 7, 2026, presents a planning method based on deliverables, clip duration, take ratio, resolution, model rate, contingency, and human review. Its three campaign examples are explicitly illustrative. Using the article’s assumptions, review accounts for 71%–76% of total cost. That is not an industry benchmark and should not be reused as one.
A companion article from the same author and date describes an illustrative production pipeline with staged review gates. Draft ideas are screened at lower resolution before expensive final renders. It also recommends treating video generation as asynchronous work, storing task IDs, making submissions idempotent, copying successful outputs to durable storage, and separating technical retries from creative rejections.
The important lesson is broader than Wan3.0: model price is only one multiplier. Your approval process determines how many times that multiplier is paid.
Build two ledgers, not one
A useful budget has a generation ledger and an approval ledger.
The generation ledger records, for each deliverable:
- number of clips;
- seconds per clip;
- generated takes per approved clip;
- resolution and rate used at each stage;
- technical retries kept separate from creative alternatives.
A basic formula is:
generation cost = Σ(clips × seconds × takes × rate per generated second)
The approval ledger records:
- minutes needed to inspect each take;
- minutes needed for final sign-off;
- loaded hourly cost of reviewers;
- rejection reason: technical failure, brief mismatch, brand issue, or subjective preference.
Then:
approval cost = review hours × loaded hourly rate
and:
cost per approved second = (generation cost + approval cost + editing cost) ÷ delivered seconds
This denominator matters. A cheap model that requires many unusable takes can cost more per approved second than a higher-priced model with a better acceptance rate.
A self-designed example
Illustrative calculation, not a benchmark or measured result: suppose a campaign needs six approved clips, each 12 seconds long. Assume five takes per clip, all generated at 720P. Using the source article’s October 6, 2026 Singapore list-price example of $0.10 per generated second:
- Generated seconds:
6 × 12 × 5 = 360 - Generation cost:
360 × $0.10 = $36.00 - Delivered seconds:
6 × 12 = 72
Now assume a reviewer spends 1.5 minutes on each of 30 takes, at a self-assigned loaded rate of $45 per hour:
- Review time:
30 × 1.5 = 45 minutes = 0.75 hour - Review cost:
0.75 × $45 = $33.75 - Simplified total:
$36.00 + $33.75 = $69.75 - Cost per approved second:
$69.75 ÷ 72 ≈ $0.97
The API rate is $0.10 in this dated vendor example, but the simplified approved-second cost is about $0.97 because generation is repeated and review is paid. This does not prove that every workflow has the same ratio.
Next, test a staged policy: four drafts per clip at 480P using the cited $0.05 rate, followed by one 720P final at $0.10.
- Draft cost:
6 × 12 × 4 × $0.05 = $14.40 - Final cost:
6 × 12 × 1 × $0.10 = $7.20 - Generation total:
$21.60 - With unchanged review assumptions:
$21.60 + $33.75 = $55.35 - Cost per approved second:
$55.35 ÷ 72 ≈ $0.77
Inference: staged resolution could reduce spend if low-resolution drafts predict creative acceptance. If they hide defects that only appear in the final render, the apparent saving may create more rework. That must be tested with your own clips.
Count the seconds you can deliver
Design gates that produce evidence
A review gate should have a written pass condition and a recorded output. “Looks good” is difficult to reproduce. “Product shape is correct, mandatory action occurs by second 6, no unapproved text appears, and the final frame leaves two seconds for the logo” can be checked by another person.
Track at least four ratios:
1. generated seconds per approved second; 2. takes per approved clip; 3. review minutes per take; 4. technical retries per submitted job.
When a ratio worsens, its category suggests the fix. More technical retries point to job handling. More creative takes point to the brief or model fit. Longer review time points to unclear criteria or poor comparison tools.
Hands-on exercise: price one real shot
Choose one 8–15 second deliverable you expect to make this month. Before generating anything, write three pass conditions. Create a small table with one row per take and these columns: resolution, generated seconds, result, rejection reason, review minutes, and approved yes/no.
After the first approved clip, compute generation cost, review cost, and cost per approved second. Mark every assumed price with its region and access date. Your checkable outcome is not a “good video”; it is a ledger in which every generated second and every rejection has a reason.
Source articles, both checked October 12, 2026:
Understand it. Then try it.
Read something elseNo sign-in needed. One like per browser, across all languages.
Original references
Use these links to verify details and learn more. A reference does not imply endorsement.
www.alibabacloud.comReference 1www.alibabacloud.comReference 2