Seedance 2.5 for ads: what it does, what a 30-second ad costs, and how to prompt it

Seedance 2.5 is ByteDance's video model that renders up to 30 seconds in one take, with sound generated alongside the picture. For ads that means a full hook, demo and close without stitching. It is also expensive: about $0.47 per second at 720p on fal, or 2,115 Orisu credits for a 30-second ad. Use it to finish proven scripts, not to find them.

This post merges our launch-week notes with what we know now that the model is generally available. It covers what changed, what it costs, a prompt format that works for ad shots, and where it fits in a testing workflow.

What is Seedance 2.5?

ByteDance announced Seedance 2.5 at the Volcano Engine FORCE Conference on June 23, 2026, and opened the API on BytePlus on July 16, 2026. Since then it has landed on fal, including a US-hosted copy, which is where most teams outside China now reach it.

It comes in three flavours: text-to-video, image-to-video and reference-to-video. On fal all three share the same settings:

The two changes that matter for ad work are the length and the reference budget. Thirty seconds holds a complete performance ad. Fifty reference files means you can hand the model product photos from several angles, frames of your actor and a voice sample instead of describing them in prose.

Why does a single 30-second take matter for ads?

Most AI video ads today are cut together from 5 to 10-second clips, and the joins are where viewers notice. A face shifts a little between shots, the lighting changes, the product grows by a few percent, or a voice track recorded separately drifts off the lips.

A single take avoids those joins. One generation holds the actor, the light and the audio from the first frame to the last, so lip sync is part of the clip rather than a repair step. That does not make every take good. It means the failure modes are inside one clip, where you can judge them, instead of in the edit.

What does Seedance 2.5 cost?

fal's model page, checked on October 6, 2026, lists roughly $0.2205 per second at 480p, $0.4730 at 720p and $1.164 at 1080p, the same for all three flavours. Here is what that becomes on Orisu's catalog, with dollars at the Pro plan rate of $0.0098 per credit:

For comparison, Seedance 2.0 at 720p is 45 credits per second, Kling 3.0 Standard with audio is 18.9 and Veo 3.1 with audio is 60. The full table, 28 settings across 16 models, is in our AI video model price index.

Two practical notes. If you leave the duration on "auto", the model picks the length and Orisu reserves credits for the longest possible clip, then settles to what was delivered. And 1080p costs about 2.5 times 720p, so render tests at 480p or 720p and only go up for the cut you will scale.

How should you prompt Seedance 2.5 for an ad?

Write the ad as shots with timings, not as a paragraph. A long model will fill an unstructured 30 seconds with whatever it likes. A shot list tells it where the hook ends and the demo begins. This is the format we use, adapted from our guide to structured prompting for AI video:

Format: vertical 9:16, handheld phone footage, natural window light, 30 seconds.
Actor: woman in her 30s from reference images 1-3, same outfit throughout.
Product: the bottle in reference images 4-6. Keep the label text exactly as shown.

0-3s  HOOK. Close-up, she holds the bottle to camera.
      She says: "I stopped buying three different serums."
3-12s DEMO. She applies two drops to the back of her hand, rubs it in,
      camera follows her hand. Quiet room tone.
12-24s PROOF. Medium shot, bathroom mirror. She talks to camera about
      how it feels after a week. Relaxed, unscripted delivery.
24-30s CLOSE. Bottle on the counter, label facing camera, she taps the cap.
      She says: "Link's below."
Audio: her voice only, no music, light room echo.

What makes this work:

1. Timecodes for every beat. The model has a plan for all 30 seconds, and the hook is forced into the first three.
2. References instead of adjectives. "The bottle in reference images 4-6" holds the product better than any description. With reference-to-video you can attach up to 30 images, so include the label straight on.
3. The exact words in quotes. Native audio follows quoted dialogue far more closely than a description of what she "talks about".
4. One instruction for audio. Say what should be heard and what should not. "No music" prevents a generic bed you would have to strip later.

Labels are still the weak point for every video model. If the text on your packaging has to be legible in the final ad, plan a product insert from a real photo for the close, and see our notes on AI product ads that keep the real label. For ready-made starting points, the Seedance 2.5 prompt hub collects prompts by use case.

Where does it fit in an ad testing workflow?

The cost table decides this. Testing ten hooks as 30-second Seedance 2.5 takes at 720p would cost 21,150 credits before you learned anything. Testing them as 4-second clips on a cheaper model costs a small fraction of that, and hooks are judged in the first two or three seconds anyway.

This is the same draft-cheap, finish-expensive pattern we recommend in which AI video model to use when. Seedance 2.5 adds a new final stage rather than replacing the stages above it.

How to run it in Orisu

Seedance 2.5 is a model option on the video node, next to Seedance 2.0, Veo 3.1, Kling 3.0 and the rest of the catalog. If you already have a UGC workflow (script, voice, actor references, product shots, video), upgrading it is a model swap on one node. The node shows the cost for your chosen length and resolution before you run, and re-running after a script change only recomputes the steps downstream of that change.

The UGC ads workflow is a good place to start if you don't have one yet. Draft the hooks on a cheap model, keep the winner, and give Seedance 2.5 the long take.