We Made a Super Bowl-Style Ad With Seedance 2.5

We wanted to find out what an AI Super Bowl commercial actually looks like when you stop treating generated video as b roll and write it like a broadcast spot: a premium 15 second film with a spectacle open, a claim, a turn and a brand card. One prompt, one 15 second generation, 720 credits of plan budget. This is the whole thing, beat by beat, with the prompts, the timings, the cost and the reasoning behind each creative decision.

That is the spot, run through two models on the same prompt. The top panel is labelled Seedance 2.5 and the bottom Seedance 2.0. Both open on a woman walking on a mirror calm ocean at sunset, both cut to a founder at a laptop, and both end on the brand card. Watching them stacked is the fastest way to see what changed between versions, and there is a fuller breakdown in our Seedance 2.5 versus Seedance 2.0 comparison.

The concept

A mock luxury commercial that reveals itself as generated.

A woman in a crimson gown walks on a mirror calm ocean at sunset and says, "They spent two million dollars on this ad." The world de-renders around her. Cut to a founder at a laptop at night, and the line, "They didn't. One prompt. No crew. No camera." Then a fast montage of the same actor in different looks and languages, then a brand card.

The joke is the structure, not a gag. The spot spends its first five seconds earning the audience's belief that it is expensive, and then spends the next ten taking that belief apart. That only works if the expensive part is genuinely convincing, which is why the opening shot got more iterations than the rest of the spot combined.

How we built an AI Super Bowl commercial in six beats

Fifteen seconds, six beats. Timings first, because the timings are the script:

Each beat below is a paragraph of one continuous prompt. Seedance 2.5 reasons about the whole shot at once, so the six beats go in as a single description of a single 15 second generation rather than six clips stitched later. If a beat is not landing, isolate it as a 4 second 480p generation at 92 credits, tighten the wording there, and fold what survives back into the full prompt.

Beat 1: spectacle, 0.0 to 2.4 seconds

Aerial drone orbit, 45 degree down angle, slowly descending, counter clockwise. A woman in a flowing crimson silk gown walks barefoot on the surface of a mirror calm ocean at golden hour, toward a low molten sun. Slow concentric ripples radiate from each step. Slow motion, anamorphic lens flare, teal and orange cinematic grade, rich blacks, gentle halation. Epic luxury commercial style, 16:9, photoreal.

Beat 2: the claim, 2.4 to 5.0 seconds

The orbit slows and settles into a low tracking shot at water level, moving with her. She turns her head toward the lens without breaking stride and says, evenly and without smiling, "They spent two million dollars on this ad." Her voice is close and dry, as if recorded in a booth rather than outdoors. Under it, only the sound of water moving against her feet. No music.

Beat 3: the reversal, 5.0 to 6.4 seconds

The world de-renders around her in one continuous move: the sun, the horizon and the water break apart into drifting geometry and then flat grey emptiness, while she stays sharp, still walking, unbothered. The camera holds on her as everything behind her leaves. The ambient ocean sound cuts to a hard digital silence in the same instant. No music.

Beat 4: the reveal, 6.4 to 9.2 seconds

Cut hard to a small dark room at night. A founder in a plain t shirt sits at a laptop at a cheap desk, one lamp, a cold window behind him, the screen lighting his face. He is watching the same footage on his laptop. He says, flatly, "They didn't. One prompt. No crew. No camera." Handheld, slight sway, natural grade, no colour styling. Room tone, a fan, a keyboard. No music.

Beat 5: the montage, 9.2 to 12.0 seconds

Fast cuts, roughly half a second each, of the same woman from beat one in different looks and different languages: a white studio backdrop, a rain slicked street at night, a bright kitchen, a rooftop at dusk. Same face throughout, different wardrobe, different light, a fragment of a different language in each. Cut on hard sound: each transition lands on the first syllable. No music, no ambience carried between cuts.

Beat 6: the brand card, 12.0 to 15.0 seconds

Cut to black, hold for a beat of full silence, then a single centred wordmark fades up on flat black with one soft horizontal light sweep across it. Nothing else in frame. The card holds still for the last two seconds. No music, no voice, no ambience.

What an AI Super Bowl commercial actually costs

Fifteen seconds at 720p, at 48 credits per second, is 720 credits, which is about $7.20 of plan credit per full render. That is the number that matters, because you will render it more than once.

Credits are the billing unit and one credit equals one cent of generation cost. The plans are Starter at $29 a month for 1,400 credits, Pro at $99 for 5,600, and Premium at $199 for 12,000. So a full 15 second 720p render is roughly half of one Starter month, about an eighth of Pro, and about a sixteenth of Premium.

Iteration costs more than the final render, as it always does. Testing single beats at 4 seconds and 480p is 92 credits a go, which is what makes the iteration affordable: you can throw away nine attempts at the opening shot for less than 900 credits and still have most of a Pro month left for the real renders.

The honest way to read this is not "a Super Bowl ad for seven dollars". A national broadcast spot is a media buy, not a production budget, and nothing here touches that. What it does mean is that the production side of a premium looking 15 second film is no longer the constraint on whether you make one.

Why the reversal works

Most generated video ads announce themselves in the first frame, and the viewer spends the whole clip auditing the pixels instead of listening. This spot uses that instinct as the mechanism.

For five seconds the audience is doing exactly what they always do, looking for the tell. Then the spot tells them, before they find it, and reframes the entire opening as a demonstration rather than a deception. The confession converts a liability into the proof.

Two craft notes made it land. The reversal is the shortest beat in the spot at 1.4 seconds, because a de-render that lingers becomes a visual effect rather than a rug pull. And the audio cut is harder than the visual one: ocean ambience to hard silence in a single frame. Since Seedance 2.5 generates audio with the video, including sound effects, ambient sound and lip synced speech at no extra cost, you can write that silence into the prompt as a deliberate beat rather than fixing it in an edit.

Why the one ordinary human shot matters

Beat 4 is the only unglamorous shot in the film. No grade, no camera move worth naming, one lamp, a person at a desk. It is the cheapest looking thing on screen and it is doing the most work.

The spectacle beats are impressive and weightless. Nothing in them is at human scale, so nothing in them is believable in the way that matters. Cutting to a normal room at normal exposure gives the claim a body to come out of. Take that beat out and the spot is a tech demo. Leave it in and it is an argument delivered by someone.

It has to be shot plainly to work. Every instinct says to grade it, move the camera, put a rim light on him. All of that would flatten the contrast between the two worlds and cost the spot its centre. We wrote "natural grade, no colour styling" into that paragraph specifically to stop the model making it beautiful.

Why the disclosure line is the flex

"One prompt. No crew. No camera." is the disclosure. It is also the strongest claim in the ad.

That is a rare piece of luck in advertising: the thing you are obliged to be transparent about is the thing you most want to boast about. It means the spot never has to carry a small grey disclaimer at the bottom of the frame, because the disclosure is the punchline, delivered on camera, in the middle of the film.

This is worth copying even when your product has nothing to do with video. If some part of how you make the thing is unusually efficient, say it in the ad in the voice of the ad, rather than in a footnote. And when generated video does need a formal disclosure, the platform rules are the place to check: both Meta for Business and TikTok for Business publish their own requirements for realistic AI generated content, and they change.

What we would change

The montage in beat 5 is the weakest beat. Half second cuts push the model hard, and identity across four rapid wardrobe and language changes was the thing we regenerated most. If we ran it again we would give the montage a full four seconds and three looks instead of four, and we would attach the woman as a reference image so the mode becomes reference to video and her face is bound to a file rather than to a description.

The other change is length. At 15 seconds the montage and the card fight each other for room. At 20 or 25 seconds, both breathe, and Seedance 2.5 generates up to 30 seconds in a single pass rather than stitching shorter clips together, which is covered in our post on 30 second Seedance 2.5 videos. The cost scales linearly: 25 seconds at 720p is 1,200 credits.

If you want to build something like this, start with the structure rather than the images. Write the beats and their lengths first, then write one paragraph of prose per beat, then generate. Our guide to turning an ad script into video covers that translation step, the Seedance 2.5 prompt guide covers the sentence level craft, and our Seedance 2.5 review has the wider assessment of where the model holds up. ByteDance's Seed research site documents what the model is designed to do.

Frequently Asked Questions

How much did this AI Super Bowl commercial cost to make?

The full 15 second spot at 720p is 720 credits per render, about $7.20 of plan credit, since Seedance 2.5 costs 48 credits per second at 720p and one credit is one cent of generation cost. Isolating a single beat at 4 seconds and 480p to test wording costs 92 credits per attempt.

Was the spot generated as one clip or stitched from six?

One clip. The six beats go in as a single continuous prompt for a single 15 second generation, because Seedance 2.5 reasons about the whole shot at once. You can still isolate a single beat at 4 seconds to tighten its wording, then fold that wording back into the full prompt.

How long can a single Seedance 2.5 generation be?

Up to 30 seconds, selectable in the editor at 4, 8, 12, 15, 20, 25 or 30 seconds. A 30 second clip is generated in a single pass rather than stitched from shorter clips. At 720p, 30 seconds costs 1,440 credits, so that length starts on the Pro plan.

Did the audio need a separate tool?

No. Seedance 2.5 generates sound effects, ambient sound and lip synced speech with the video, and it is included in the price with no separate toggle. The hard silence after the reversal was written into the prompt as a deliberate beat.

Why does the spot admit it is AI generated?

Because the admission is the strongest line in the ad. It converts the audience's instinct to look for the tell into the proof of the claim, and it means the film never needs a small grey disclaimer, since the disclosure is delivered on camera as the punchline.

Can I make something like this without a video background?

Yes. The hard part is structure, not craft: write the beats and their lengths first, then one paragraph of prose per beat describing subject, camera, light and sound. Test at 480p at 23 credits per second and only render the approved version at 720p.