Research-backed calculator guide

AI Video Retry Budget: Price Failed Takes Before the Final Cut

Use an AI video generation retry cost calculator to price attempts, accepted shots, and finished seconds without hiding failed takes or mixing credits with dollars.

The calculator model: attempts, acceptance, and usable footage

Treat each shot as a small production experiment. You submit a prompt and references, receive an output or failure, review it against a brief, and either accept it, mine part of it, or reject it. The calculator does not need to guess why a take failed; it needs to preserve the cost of the attempt and make the acceptance rule explicit.

Use these fields:

1. Kept shots. The number of shots that must pass your stated brief.
2. Attempts per shot. The average number of charged generations used to obtain one accepted shot. If you track acceptance rate instead, attempts per accepted shot is approximately 1 ÷ keep rate when the rate is measured on comparable jobs.
3. Generated seconds per attempt. The requested or billed output duration, not necessarily the seconds that survive editing.
4. Price basis. Dollars per second, dollars per output, credits per second, GPU time, or a subscription allocation.
5. Scope. Whether reference-image creation, input video, audio, editing, upscaling, storage, and human review are included.
6. Acceptance rule. The observable conditions that make a shot usable, such as duration, framing, character identity, readable product text, motion, and sound.

The keep rate is useful only when its denominator is honest. If 30 of 90 attempts become accepted shots, the observed keep rate is 33.3% and the observed attempts per accepted shot is 3. That does not prove the next project will also require exactly three attempts. It is a planning assumption that should be replaced by your own ledger as soon as you have comparable data.

A provider quote also needs a date and configuration. For example, Runway’s developer pricing page says credits can be purchased for $0.01 per credit and lists Gen-4.5 at 12 credits per second, while Gen-4 Turbo is listed at 5 credits per second. Those are dated vendor rates for the documented developer service, not a universal market price and not a promise that a consumer plan uses the same meter.^[1]

Other services use different meters. Replicate explains that public models may be billed by hardware time or by input and output, and its pricing page lists Wan 2.1 image-to-video examples at $0.09 or $0.25 per second depending on output resolution.^[2] A calculator can compare those rates, but it should retain the provider, model, resolution, and billing unit beside every row.

Google’s pricing documentation is another reminder to avoid stale per-clip assumptions. Its Gemini Omni Flash section describes video output billing as token-based, at 5,792 tokens per second of 720p video, with an approximate effective price of $0.10 per second under the displayed standard example.^[3] The same page can change as models and tiers change. Check the current official pricing before committing a production budget.

Worked cost example: 24 accepted shots with three attempts each

Suppose a product video needs 24 accepted shots. Each generation is five seconds long. Your expected review history suggests three attempts per accepted shot. You choose a model whose documented rate is $0.60 per five-second attempt. That could be a direct quote, a converted credit rate, or an internal planning input; label which one it is in your own worksheet.

The arithmetic is:

• Kept shots: 24
• Attempts per shot: 3
• Total attempts: 24 × 3 = 72
• Cost per attempt: $0.60
• Raw generation estimate: 72 × $0.60 = $43.20
• Cost per accepted shot: $43.20 ÷ 24 = $1.80
• Finished footage if all 24 shots contribute five seconds: 24 × 5 = 120 seconds, or 2 minutes
• Cost per finished minute: $43.20 ÷ 2 = $21.60

This is a generation-only estimate. It excludes editing time, music or voice licensing, storage, captions, delivery exports, taxes, and any reference assets unless you add them as separate rows. If only four seconds of each accepted five-second clip survives the edit, the finished runtime is 96 seconds, not 120. The cost per finished minute then becomes $43.20 ÷ 1.6 = $27.00.

The table makes the calculation inspectable:

Now run sensitivity instead of presenting $43.20 as a promise. With 24 shots and a $0.60 attempt price, two attempts per shot produce $28.80, three produce $43.20, and five produce $72.00. The difference is driven by the acceptance assumption, not by a hidden mathematical trick. If the output duration or resolution changes the provider price, recalculate the attempt price rather than reusing $0.60.

A reserve can be added transparently. A 10% planning reserve on the expected $43.20 is $4.32, producing a budget envelope of $47.52. Call that a reserve estimate. It is not a claim that failures will stop at 10%, and it should not be used to conceal a known retry multiplier.

How to build an audit-friendly retry ledger

Record one row per operation or one row per shot attempt, depending on the scale of the project. At minimum, keep the date, service, account or plan, model, mode, duration, resolution, input references, displayed price, actual debit, output file, review status, and rejection reason. A small ledger is more useful than a polished average that cannot be reconstructed.

Separate these statuses:

• Submitted: the request started or was queued; do not call it accepted.
• Generated: a file exists; review is still required.
• Accepted shot: the file meets the shot-level brief.
• Accepted seconds: the seconds that actually enter the edit.
• Assembled sequence: clips have been combined; sequence-level checks may still fail.
• Approved delivery: the final export passed its own audio, text, timing, and format checks.

This distinction matters for credit systems. Replicate’s documentation says that many public models are billed for the time needed to process a request, while other models use input/output pricing.^[2] A model page can therefore provide a useful current estimate, but your ledger should retain the actual billing basis and the exact version or settings. If a private deployment keeps a hardware instance online while idle, that idle cost belongs in infrastructure scope rather than being silently folded into per-clip generation cost.

Likewise, a failed job is not automatically free or automatically charged across every provider. Mark the debit as verified, reversed, pending, or unknown. Do not put zero in the sheet just because the interface did not show a final balance. If a provider gives a refund, record the original debit and the reversal separately, then calculate the net only after reconciliation.

Pricing changes and model-selection checks

Model comparison pages become stale quickly. Google’s video guide states that the Veo 3.1 Preview models in the Gemini API were shut down on October 22, 2026, and directs developers toward Gemini Omni Flash or other available channels.^[4] That is a concrete reason to attach a checked date to every model row. A calculator using an old model name can produce neat arithmetic for a route that is no longer available.

Use the Model pricing comparison route to organize a provider comparison, but verify each model’s current official documentation before purchasing credits. For token-metered or mixed workloads, the AI token cost calculator can help separate input and output units from the video ledger. If the workflow also includes ordinary language-model planning, the LLM cost calculator is a better place to price that text portion than hiding it inside a video attempt.

Do not select the cheapest nominal rate without checking acceptance. A low-cost model with five attempts per accepted shot can exceed a higher-cost model that usually passes in two attempts. The right comparison is often:

effective accepted-shot cost = attempts per accepted shot × cost per attempt.

That comparison still does not settle quality, latency, queueing, licensing, safety review, or editing time. It only keeps the generation arithmetic honest.

What this calculator excludes

The retry budget is not a full production budget. Unless you add them explicitly, it excludes prompt development time, storyboard work, reference-image generation, input-video charges, storage, bandwidth, editing, color work, sound design, voiceover, music, captions, stock assets, contractor time, taxes, and failed payment or account operations.

It also does not infer an acceptance rate from a marketing claim. A vendor’s example, a community report, or a single short film can inform a scenario, but it is not your measured production yield. Label third-party observations as context and label your own future case as an estimate until the ledger is complete.

Finally, a cost per finished minute is not always comparable across projects. A documentary-style sequence, a product close-up, and a character scene can have different shot counts, reference requirements, motion difficulty, and review standards. Use the metric to compare scenarios with similar definitions, not to imply that one universal number describes AI video production.

FAQ

How do I calculate failed AI video generation cost?

Add the charged cost of all attempts tied to the task, including rejected outputs, then divide by the number of accepted shots or accepted seconds. In formula form, attempt count × cost per attempt gives raw generation cost. Raw generation cost ÷ accepted shots gives cost per accepted shot. If no shot passes the brief, report the spend and an accepted count of zero rather than inventing a usable-output price.

What is a reasonable retry multiplier for AI video?

There is no provider-neutral default that applies to every prompt, model, and review standard. Use a low, expected, and high scenario, such as 1.5, 3, and 5 attempts per shot, only when you label them as planning assumptions. Replace those values with your own observed attempts per accepted shot once your ledger has enough comparable rows.

Should I use keep rate or attempts per shot?

Use whichever you can measure consistently. If 30 of 90 comparable attempts are accepted, the observed keep rate is 33.3% and the observed attempts per accepted shot is 3. Keep rate is intuitive for yield reporting; attempts per accepted shot is usually easier to multiply into a budget. Neither one predicts a new creative brief without uncertainty.

Do credits equal dollars across AI video platforms?

No. Credits are provider-specific accounting units. Runway’s developer documentation currently describes a $0.01 purchase price per credit and model-specific credits-per-second rates, but another service may bill by hardware time, input/output, or a different credit bundle.^[1] Convert only within the same provider, account arrangement, model, and date, and keep credits and cash as separate columns.

Should editing and audio be in the retry calculator?

Keep generation retries separate from finishing unless your question is explicitly all-in delivery cost. Add editing, audio, upscaling, storage, and labor as named rows or a second scenario. This prevents a $43.20 generation estimate from being mistaken for a finished, approved video budget.

What if a five-second clip only contributes three seconds to the edit?

Count five seconds as the generated duration for a per-second charge, but count only three accepted seconds in the finished-runtime denominator if that is all you use. This distinction raises cost per finished minute and prevents a partly used clip from inflating the apparent yield.

How often should I refresh model prices?

Refresh before a production commitment and whenever you change provider, model, resolution, duration, plan, or billing route. Record the checked date beside the rate. Google’s documentation, for example, records a dated shutdown for Veo 3.1 Preview in the Gemini API, illustrating why a model label alone is not durable evidence.^[4]