Research-backed calculator guide

AI Video Cost per Approved Minute: Include Retries, Rendering and Review

Calculate AI video cost per approved minute by separating paid generations from retries, rendering, review, editing, storage, and delivery.

What “approved minute” measures

An approved minute is the runtime that survives your acceptance rule. The rule might require usable motion, continuity, brand compliance, readable text, correct dialogue, safety review, and a final export. It does not mean one minute of files returned by an API.

This distinction matters because video models usually produce short clips. A 60-second deliverable may be assembled from twelve five-second shots, twenty-four 2.5-second cuts, or a mixture. Count the shots that remain in the edit, then estimate how many paid attempts were needed for those shots. A rejected creative take can be billable even when a technical failure may be refunded. Treat those as separate events in your ledger.

There is no universal provider billing unit. fal says its Model APIs charge by the successful output, with video models using either a per-second or flat per-video unit, and says server errors and queue waiting are not charged ^[2]. Replicate explains that many public models are billed by hardware time, while some model pages use input or output pricing; its listed Wan examples show output-video prices per second ^[3]. Those are different vendor meters, so normalize them to dollars per paid attempt before comparing them.

Google’s Gemini Omni Flash documentation is another useful warning against assuming every route is priced per video second. The page lists $17.50 per million output tokens for video and says billing uses 5,792 tokens per second of 720p video, an effective price of about $0.10 per second under that standard pricing ^[4]. That is a derived effective rate, not permission to treat all models or resolutions as equivalent. Record the exact model, resolution, audio setting, minimum charge, and input treatment.

The approved-minute formula

Use the following sequence for each model or workflow lane:

1. Approved seconds: approved minutes × 60.
2. Kept clips: approved seconds ÷ average seconds kept per clip. Round up if your workflow cannot generate a fraction of a clip.
3. Paid attempts: kept clips × retry multiplier. A multiplier of 1.0 means one paid attempt per kept clip; 3.0 means three on average.
4. Generation cost: translate the provider meter into dollars per attempt, then multiply by paid attempts. For a per-second route, this is paid generated seconds × rate per second. For a flat per-video route, use the number of attempts × price per video.
5. Finishing lines: add rendering, audio, review, editing, storage, delivery, and any project allocation you intend to recover.
6. Unit cost: divide the complete numerator by approved minutes.

If the provider bills by output seconds and every attempt has the same clip duration, a compact version is:

Generation cost = approved minutes × 60 × retry multiplier × vendor rate per generated second.

That shortcut breaks when there are minimum charges, different draft and final resolutions, input-video charges, reference-image fees, or a mix of models. In those cases, calculate each lane separately. For example, you might draft background shots on a lower-rate model and reserve a premium model for hero shots. The approved-minute denominator stays common, while the numerator has several rows.

Review can be calculated independently:

Review cost = paid attempts × review seconds per attempt ÷ 3,600 × reviewer hourly rate.

This catches a cost that a flat percentage can hide. If a model has a poor keep rate, a reviewer sees more clips even though the approved runtime has not changed. Editing is usually easier to estimate as editor hours × loaded hourly rate, but use a separate line if long-form pieces require multiple passes. Keep storage and delivery separate when your retention period, CDN, or export volume is material.

Worked cost example

Assume a US-dollar planning sheet checked on 2026-11-02 with these inputs:

• Deliverable: 1.0 approved minute.
• Structure: 12 kept shots, each 5 seconds, for 60 approved seconds.
• Retry multiplier: 3.0 paid attempts per kept shot.
• Model route: Runway Gen-4 Turbo, 5 credits per generated second and $0.01 per credit from the published developer pricing page ^[1].
• Rendering: internal estimate of $0.02 per generated second, not a Runway quote.
• Review: 45 seconds per attempt at an internal loaded reviewer rate of $40 per hour.
• Editing: 1.5 hours at an internal loaded editor rate of $40 per hour.
• Storage and delivery: $2 planning allowance, not a provider quote.

The 12 kept shots require 36 paid attempts. At five seconds per attempt, that is 180 generated seconds. The vendor line is 180 × 5 × $0.01 = $9.00. Rendering is 180 × $0.02 = $3.60. Review is 36 × 45 ÷ 3,600 × $40 = $18. Editing is 1.5 × $40 = $60. Adding $2 for storage and delivery gives $92.60. Because the approved runtime is exactly one minute, the fully loaded cost is also $92.60 per approved minute.

The important result is not the specific dollar amount. It is the separation between a documented vendor meter and internal operating assumptions. If editing is already covered by salaried staff, remove or allocate it differently. If rendering is included in a platform subscription, set that row to zero but note the subscription allocation elsewhere. If your acceptance rate is better than 1-in-3, the result falls quickly.

Sensitivity to retries

Keeping every other assumption fixed, the same one-minute project changes as follows. These are original estimates using the stated internal labor and finishing assumptions, not provider forecasts.

The edit row remains $60 in this simple sensitivity table. In a real project, five attempts may also increase logging, conforming, director review, and revision time. Capture those effects instead of assuming retry cost is only a provider charge.

How to compare providers fairly

Start with the exact route, not a headline model name. On Runway’s page, Gen-4.5 is listed at 12 credits per second and Gen-4 Turbo at 5 credits per second, while some other listed models have audio, reference-video, minimum-charge, or input-video rules ^[1]. Convert the relevant row to dollars and preserve those conditions in your worksheet.

For fal, record whether your selected endpoint bills per output second or per video. Its documentation says successful outputs are charged and server errors are free ^[2]. That means a technical failure policy can affect expected cost, but a creative rejection after a successful generation remains a workflow event you must model yourself. Do not apply a universal “failure rate” without defining which failures are refunded.

For Replicate, distinguish public model billing from private deployment billing. Its pricing page says many public models are billed by runtime on the hardware used, while private models can incur setup, idle, and active time ^[3]. A per-second output example is not interchangeable with a GPU-hour route. If you self-host or use a dedicated deployment, add utilization, warm-up, idle, and operator costs to the numerator.

For token-metered video, calculate the effective rate only after confirming the provider’s token conversion. Google’s Gemini Omni Flash page expresses video output in tokens and gives a 720p seconds-to-token relationship ^[4]. If you switch resolution, model tier, or output behavior, re-check the conversion. A rate shown as “per second” may be a convenience calculation rather than the underlying meter.

For a broader provider comparison, use the site’s Model pricing comparison calculator to normalize rate, unit, and duration. The same discipline applies to adjacent token or multimodal work: the AI token cost calculator is useful when a video workflow also invokes language or vision models, and the LLM cost calculator can keep non-video orchestration out of the video row rather than silently mixing it in.

What to include and what to exclude

Include any cost that you promise to recover in the price of the approved minute. Generation and retries are the minimum. Rendering belongs in the numerator when you pay for a render service or allocate compute to encoding. Review belongs there when a person checks every attempt or every candidate. Editing belongs there when the deliverable includes assembly, captions, color work, sound mix, or revisions. Audio belongs there when voice, music, or effects are purchased separately. Storage and delivery belong there when retention, export, or transfer is not negligible.

Exclude costs deliberately, not accidentally. You may exclude script development, creative direction, paid stock, taxes, platform subscription allocation, or client management from a narrow “generation-only” number. Label that number as generation-only and show a second fully loaded figure. Do not call a generation-only figure a production cost.

Keep vendor quotes and estimates visually distinct. A provider documentation row is a dated published rate. Your reviewer wage, render allowance, and edit hours are estimates. If a provider says pricing may change, attach the checked date to the worksheet and re-run the calculation before a new project or quote. Prices can also vary by account, region, contract, tax, plan, or volume discount.

FAQ

Is AI video cost per approved minute the same as cost per generated minute?

No. Cost per generated minute divides spend by seconds returned from the model. Cost per approved minute divides spend by the runtime that passes review. Rejected takes make the approved-minute number higher, sometimes substantially higher, even when the provider’s per-second rate does not change.

What retry multiplier should I use?

Use measured attempts per approved shot when you have project history. Before you have data, model at least a low, expected, and high case such as 1.5×, 3.0×, and 5.0×. The right value depends on prompt control, character continuity, dialogue, motion complexity, and your acceptance standard. Treat it as a planning assumption, then replace it with logged attempts and approvals.

Should human review be charged for rejected clips?

If a reviewer watches or checks the rejected clip, yes. Review time follows paid attempts, not only approved clips. A technical failure that never produces a viewable file may take little review time, while a creative rejection can consume the same review seconds as an approved take. Track those paths separately.

Do rendering and editing belong in the approved-minute price?

They belong in the fully loaded price when the buyer expects a finished, exported video. You can publish a generation-only number as a useful diagnostic, but it should not be presented as the cost of a finished minute. Show the assumptions for render rate, editor hours, revision rounds, audio, and export format so another person can reproduce the total.

How often should I refresh the calculation?

Refresh whenever the model, plan, resolution, audio mode, or billing unit changes, and before using the number in a new commercial quote. This article’s provider observations were checked on 2026-11-02. A dated rate is evidence for that date, not a promise about a future invoice. Keep the original rate and the refreshed rate in your ledger so variance is explainable.