What AI-generated video ads actually cost

Nobody publishes this. Every AI video tool quotes a subscription price, which tells you what they charge and nothing about what the thing costs to make. That gap is where all the interesting questions live: whether the pricing survives contact with heavy usage, whether a “credit” means anything, and what happens to your bill when a campaign actually works.

We have generated 51 of these ads for real apps. Here is what they cost, taken from the vendor’s own published rates and our own per-call logging.

A 25-second ad costs about $2.57

The generation model underneath is BytePlus’s Dreamina Seedance 2.5, which bills per output token rather than per video. At 480p that works out to $0.514 for five seconds, or $0.1028 a second, published on their own pricing page. Resolution is the single biggest lever:

480p sounds low until you remember where the ad is watched: a 9:16 clip inside a feed on a phone, usually at partial brightness with the sound off. The resolution that matters for performance is the one the platform re-encodes to, not the one you uploaded.

Duration is linear, and that is a trap

Because billing is per token and tokens scale with output length, cost is almost exactly linear in duration. A 30-second ad costs six times a 5-second one. There is no bulk discount for a longer clip.

The practical consequence: if you are testing hooks, test them at five seconds. Six short cuts cost the same as one long one, and hook testing is where most of the learning is anyway.

Sending a video in costs less than not sending one

This one is counterintuitive. Seedance bills at two different rates depending on whether the request includes a video input:

The cheaper rate applies when you condition generation on footage you supply. But the total is not automatically lower, because cost there scales with the input video’s duration as well: a 5-second 480p output costs $0.553 with a 2–4 second input and $2.152 with a 30-second one. Feed the model a long clip and you pay for it.

Prepaid token packs are not a discount

BytePlus sells prepaid resource packs — 5 million tokens for $32, which reads as $6.40 per million against a $10.70 pay-as-you-go rate. A 40% saving, apparently.

It is not. Packs deduct at a different ratio depending on request type. For input without video — which is most UGC generation — the documented ratio is 1 : 1.671875. Every token you consume draws 1.671875 tokens from the pack:

$6.40 × 1.671875 = $10.70, to the cent. The ratio is calibrated so the pack costs exactly the same as paying as you go. For requests that do include video the ratio is 1:1 — also identical. There is no configuration in which a pack is cheaper.

What a pack does carry is a 90-day expiry on unused tokens, no refunds, and lock-in to one model version. If you are buying them expecting a discount, you are accepting three constraints for nothing. They are a payment method, not a saving.

What this means for a real budget

Creative testing is a volume problem — industry estimates put the share of creatives that ever become winners at roughly 4–8%. That number is the whole argument for generated video, and it makes the per-clip cost the only figure that matters:

For comparison, a single commissioned UGC video from a creator typically runs into the low hundreds of dollars — meaning the entire monthly test volume above costs less than one of them. That is the actual case for generated creative, and it does not depend on the generated clip being better than the commissioned one. It only depends on being able to afford to be wrong.

The costs nobody quotes

Generation is not the whole bill. Two smaller line items are worth knowing about because they are invisible until you look:

• Quality checks. We run a vision model over sampled frames of every clip to verify it matches the format that was ordered. About $0.01 a clip, and it is the difference between catching a broken ad and sending one.
• Transcription for captions. Roughly $0.0025 for a 25-second clip with word-level timings. Effectively free, and it is also the only reliable way to detect the model saying your product’s name wrong.

Both together are under half a percent of the generation cost. Any vendor not doing some version of them is saving a rounding error and shipping you the consequences.

Method

Per-clip costs are calculated from BytePlus’s published rates for dreamina-seedance-2-5 and cross-checked against our own per-call logging across 51 generated ads. Deduction ratios are from the resource pack documentation. Creative win-rate figures are industry estimates, not our measurement. Prices were current in September 2026 and vendors change them.