Research-backed calculator guide

Cost per Finished AI Ad: Compare One Hero Cut with a Variant Portfolio

Estimate AI video campaign cost per finished ad by separating generation, retries, editing, rights, review, and the marginal cost of extra formats.

What “finished” means in this calculation

Define the output before entering a price. In this guide, a finished ad is an exported asset that has passed the agreed creative, product-fidelity, copy, audio, aspect-ratio, rights, and compliance checks. It is ready to upload to a paid-media platform. A draft, an unedited clip, or a generation that still needs a product-label correction is not finished.

That definition prevents a common denominator error. A tool may charge per second, per credit, per render, or per subscription. Those are input meters. An advertiser ultimately needs an approved output meter. Replicate documents both time-based hardware billing and model-specific input/output billing, which means the same workflow can have different cost behavior depending on the model and deployment path. ^[3]

The denominator should also say whether one ad means one placement or one creative idea. My default is one approved creative in one specified format. If the same concept is exported to 9:16 and 1:1, count two finished deliverables when both files are independently reviewed and scheduled. If your production team treats the second orientation as a cheap derivative, show that marginal cost separately rather than hiding it.

Build the campaign cost stack

Use six lines so a low generation rate does not hide the work around it.

1. Fixed setup. Include brand-context preparation, templates, project setup, and any monthly subscription allocation. Allocate only the share used by this campaign.
2. Generation attempts. Price every clip request that consumes credits or compute, including rejected takes and planned retries.
3. Assembly and versioning. Include editing, captions, audio mix, exports, and cutdowns. A second orientation may reuse the timeline, but it still needs a rendered and checked output.
4. Voice, music, and rights. Include paid voice generation, music licensing, stock, likeness permissions, and paid-media usage terms where relevant.
5. Review and corrections. Price product-fidelity checks, claims review, brand review, accessibility checks, and the time spent routing approvals.
6. Contingency. Use a clearly labeled allowance for uncertain retry or revision volume. Do not disguise it as a measured failure rate.

For API work, keep the vendor meter visible. Runway Dev states that credits cost $0.01 each and publishes model-specific rates such as 5 credits per second for gen4_turbo; it also documents higher rates for premium models, audio, resolution, and some professional formats. ^[1] Those are vendor inputs, not a finished-ad quote. A six-second clip at 5 credits per second is $0.30 before any retry, edit, or approval work.

Google’s Gemini API pricing page similarly separates free, paid, and enterprise access and notes that production features such as Batch API can have different economics from standard usage. ^[2] Treat the model, tier, batch mode, resolution, audio setting, and date as part of the input record. Pricing pages change, so label time-sensitive values with the date checked.

Worked cost example: ten finished ads

The following scenario is intentionally inspectable. It is a planning model for a U.S.-English campaign, not a quote from an agency or a guarantee of output quality.

The table uses $27 as the vendor-priced generation line. The other amounts are explicit estimates that you should replace with your own rates, invoices, or time logs. If you exclude labor, the platform-only figure is $2.70 per approved ad, but that is not the campaign’s finished-ad cost. If three attempts per keeper is too conservative for your workflow, measure your own ratio for a week and replace it. If the actual ratio is five attempts, the vendor line becomes $45 rather than $27, adding $1.80 per finished ad in this small example.

The same math can be written as:

C_finished = (F + (A × clips × seconds × attempts × credits_per_second × credit_price) + E + R + Q) ÷ A

Here, A is approved ads, F is allocated fixed setup, E is editing and versioning, R is rights and audio, and Q is review. Keep units consistent: seconds with credits per second, dollars with dollars, and hours with a loaded hourly rate.

Hero cut versus variant portfolio

A hero cut often has one concept, more art direction, and a larger polish budget. A portfolio has several hooks or executions designed to learn which message works. Compare them using a common base brief.

Suppose a hero production has $29 of setup, $27 of generation, $180 of editing, $60 of review, and $40 of rights. Its estimate is $336 for one approved ad. A ten-ad portfolio might reuse the $29 setup, use $27 of generation per the worked assumptions, and add $240 editing, $120 review, and $40 rights for a total of $466. The portfolio is more money in absolute terms, but its estimated average is $46.60 per approved ad. The hero is $336 per approved ad. Neither number says which creative will perform better; it only describes production economics.

The portfolio advantage appears when a campaign can reuse approved brand context and templates. The fixed setup is no longer charged in full to every version. The trade-off is review capacity. Ten ads can require more claim checks, product checks, subtitles, and exports than one hero. Treat approval hours as a scaling line, not as free overhead.

Price the second orientation and cutdowns

Do not call an extra format free merely because the model can reuse the prompt. Estimate the marginal work.

For the ten-ad example, suppose the second orientation reuses the selected footage and brand template. Generation is $0 because no new model clip is requested. Editing is 1.5 hours at $60, caption reflow is 0.5 hours at $60, and review is 1 hour at $60. The second orientation adds $180, or $18 per additional finished format across ten ads. If every ad instead needs a new crop, a new voice track, or a new end card, increase the line. If the model must regenerate shots for safe framing, add generation attempts too.

A 6-second cutdown is different again. It may reuse the hook but need a new opening, CTA timing, subtitle layout, and compliance review. The right question is not “does the tool export it?” but “what work must pass before it can receive media spend?”

For comparing models, use the site’s Model pricing comparison to keep rate inputs in one place, then carry the chosen model and date into your campaign worksheet. If text or script generation is part of the workflow, a token cost calculator can be a separate line rather than an invisible assumption. For broader recurring model spend, use the LLM cost calculator, but do not mix text-token costs into video generation without naming the unit.

How to reduce cost without falsifying the denominator

First, lock the brief and reference assets before spending video credits. Images, product references, and approved copy can reduce avoidable regeneration. Second, test a lower-cost model at draft quality, then reserve a premium model for finalists. Third, batch common setup work: brand context, caption style, end cards, and export presets. Fourth, record every generation and its disposition: shipped, rejected for quality, rejected for product fidelity, or failed technically.

Do not use a vendor’s cheapest per-second rate as proof that a campaign will be cheap. Replicate says public models can be billed by runtime or by input/output, and that model pages provide estimates. ^[3] The billing meter still needs to be multiplied by your actual request shape. A 15-second output, three reference images, audio, and a failed retry are not equivalent to a single five-second text-to-video request.

Likewise, do not treat an observed result from one provider as a universal benchmark. The official Runway page includes model and output-format differences, while Google’s page distinguishes service tiers and changes over time. ^{[1] [2]} Use dated vendor inputs, then measure your own usable-take rate and labor hours.

What to exclude from this unit cost

Keep media spend outside production cost. This guide estimates the cost to create an approved asset, not the cost to buy impressions, clicks, or conversions. Do not put expected ROAS, revenue, ranking, or first-week performance into the denominator. Those are outcome measures and need a separate experiment design.

Also exclude unrelated company overhead unless you are explicitly building a fully loaded finance view. If you include salaries, allocate a documented share of loaded time. If you include a subscription, allocate it once; do not also count the same subscription inside every per-ad line. Taxes, currency conversion, enterprise minimums, and storage may matter, but label them rather than silently folding them into a vendor rate.

FAQ

What is the best unit for AI video campaign cost?

Use cost per approved, media-ready ad at a defined scope. State the length, aspect ratio, number of clips, audio treatment, rights, revision allowance, and review standard. Keep raw cost per second as a diagnostic input, not as the campaign headline.

How many AI video variants should a campaign budget include?

There is no universal count. Start with the number needed to test the distinct hypotheses in your brief, then add required placement or language versions. A portfolio should be large enough to answer a marketing question but small enough that each ad receives a real product and compliance review. Model low, expected, and high variant counts rather than assuming one “standard” slate.

Does a failed AI video generation count as a finished-ad cost?

It does not count in the finished-ad denominator, but its compute or credit cost belongs in the numerator when the vendor charges for it. Track technical failures separately from quality rejects. The distinction helps you decide whether to change a model, improve prompts, or increase the retry allowance.

Is a second orientation free if the first ad already exists?

Usually not. It may have little or no new generation cost, but it can still require reframing, caption reflow, audio changes, new end cards, export, and review. Estimate those marginal hours and divide by the number of additional approved files. If new shots are required, include their generation cost too.

Should internal labor be included in cost per finished ad?

Yes when the purpose is a make-versus-buy or budget decision. Use a loaded hourly rate and record the hours spent briefing, prompting, selecting, editing, correcting, and approving. You can publish a platform-only subtotal as a secondary view, but calling it the finished-ad cost understates the workflow.

How should I compare a hero ad with ten AI ad creatives?

Use the same definition of finished and compare both total cost and average cost per approved ad. The hero may carry more concept and polish work, while the portfolio spreads setup across variants. Then keep media testing costs separate so a cheaper production route is not mistaken for a better-performing creative.