A generator can look inexpensive in a price table and expensive in actual production. Credits per video second measure only successful or attempted output duration under a service’s billing rules. A studio needs the cost per accepted second after iterations, upscales, editorial repair, localization and failed jobs.
Start with a consistent unit
For each platform, capture the displayed unit exactly: Adobe generative credits, Runway credits, HeyGen credits or currency-priced API seconds. Never convert one company’s credits to another company’s credits by matching their numbers. Runway lists Gen-4.5 at twelve credits per output second. Adobe’s September 2026 rate card lists its own video at fifty credits per second for 720p and twenty at 540p. Both are source-backed costs in different currencies of account, not competing dollar prices.
Build a ledger with job_id, provider, model, version, plan, requested_seconds, delivered_seconds, charge_units, count, render_status, retries, accepted_seconds and editor_minutes. If a service charges on submission, record failed charges separately from any automatically refunded balance. If it refunds failed jobs, distinguish a promise in the rate card from an observed refund on the invoice.
Estimate usable output
Suppose a sequence needs sixty accepted seconds, the chosen mode averages six billed seconds per attempt and only one in three clips is usable. That implies thirty attempts and 180 generated seconds before assembly, not sixty. At twelve Runway credits per second, that hypothetical workflow would require 2,160 Runway credits before export, upscale and human labor. The one-in-three acceptance rate here is an illustration, not a tested model outcome. Replace it with local job logs when available.
One clip may be visually acceptable yet have an unusable audio phrase. Another may work after cropping and grading. Set acceptance categories in advance: fully accepted, accepted with edit, partially usable and rejected. Count partially usable seconds honestly, and measure staff time. The real production metric is often total cost divided by accepted, delivered minutes rather than token or credit price alone.
Queue and delivery risk
Priority tiers, free queues, delayed generations and concurrent-job limits affect whether a project arrives by its deadline. Adobe documents a special delayed queue for certain Premium users after credit depletion, with up to twenty pending video generations. That is not a reliable guarantee about when the queue will finish. The budget should include a scheduling buffer and an approved alternative route, not invented completion-time claims.
Connect to procurement
AI 10 owns centralized currency prices, active plans, tax and API unit prices. AI 04’s per-second formulas should link to the appropriate dated price record, not duplicate volatile dollar amounts in every tool article. Refresh source URLs, model entitlement, region and the ‘verified at’ date before approving major spend.
Worked production example
A hypothetical sixty-second sequence needs ten accepted six-second shots. With a one-in-three illustrative acceptance rate, the team may request thirty clips; that does not mean any particular provider has a thirty-three-percent success rate. The ledger should record every submitted generation, actual charge, usable seconds and editing time. After the first five attempts, compare the real rejection reasons with the storyboard. If almost all failures involve hands touching objects, redesign those shots or choose filmed material instead of allowing the retry budget to grow without limit.