The spec that actually matters

ByteDance released Seedance 2.5 on July 31, 2026, first on Jimeng AI and the Pro tier of Doubao, with API access on Volcano Engine's Ark platform reported as opening soon. Most of the coverage led with benchmark scores. Those aren't the interesting part.

The interesting part is one number: 30 seconds in a single generation, no stitching.

Every AI video model before this made you a shot machine. You'd get five or ten seconds, then another five or ten, then you'd sit in an editor trying to make cut two look like it came from the same universe as cut one. The lighting drifted. The jacket changed shade. Your subject's face aged three years between shots. Most of the work in an AI video pipeline wasn't generating, it was reconciling.

A native 30-second take collapses that. Thirty seconds isn't an arbitrary length. It's the length of a pre-roll ad, a product teaser, a founder story, a category explainer. It's the unit marketers actually ship.

Why one-take beats stitched, for ads specifically

Stitching has a tell. Even when you nail the continuity, the audience feels the seam, a micro-reset in lighting or grain at each cut that reads as "assembled" instead of "filmed." On a brand asset running against real production, that tell is the difference between an ad and a demo of an AI tool.

Continuity inside one generation also means the model holds the same subject through a wide shot, a push-in, and a reaction beat. You can direct a small arc instead of composing a slideshow. For a product ad that's the whole game: setup, problem, product, payoff, in one breath.

Practical version: if you've been generating six clips and spending an hour in an editor per ad, most of that hour goes away. Not because generation got smarter, but because the reconciliation step disappeared.

Here's the same prompt through both generations, 2.0 on top and 2.5 below. Watch the environment detail and how the subject holds together as the camera moves.

The 50-reference input is the brand-safety feature

Seedance 2.5 accepts up to 50 reference assets in one input, reported as up to 30 images, 10 videos, and 10 audio clips, alongside scripts and style guides. The previous generation capped out around a dozen files.

Read that as a brand consistency lever, not a novelty. It means you can feed the model:

Your actual product from six angles, so it renders your packaging and not a generic lookalike.
Your existing brand film, so the grade and camera language carry over.
Your spokesperson or UGC creator, across enough frames that the face stays the same face.
Your voice reference, so the read matches your other assets.

Earlier models made "on-brand" a prompt-engineering problem. You'd write increasingly desperate paragraphs about your color palette and hope. Fifty references turns it into a supply problem instead, which is a much better problem to have. You either have clean reference material or you don't, and that's fixable.

If you've never built a reference library, this is the reason to start. Shoot your product properly once. Pull twenty clean frames of your best-performing creator. Save your brand film's grade. That library outlives whichever model is winning this quarter.

Timestamp editing and green-screen control

Two more features matter for anyone doing repeatable ad production rather than one-off art.

Timestamp-based editing means you can address a specific moment in the clip instead of rerolling the whole thing. That's the difference between "the third second is wrong so let's regenerate and pray" and "fix the third second." It turns video generation into something closer to editing, where a client note costs you a small correction rather than a fresh lottery ticket.

Green-screen plates and 3D blockouts let you hand the model a camera move instead of describing one. ByteDance is calling the white-model blockout a first for video generation. Whether or not that claim survives scrutiny, the direction is right: prompts are a bad interface for camera language, and previz is a good one. If you've ever written "slow dolly in, 35mm, slight handheld" three different ways and gotten three unrelated moves, you already know why.

Where it sits against Veo and Kling

The field just got smaller. OpenAI discontinued Sora: the app and web experience shut down on April 26, 2026, and the API stops on September 24, 2026. If Sora is still in your pipeline, you're working against a clock, and any comparison chart still recommending it was written before the announcement.

Among what's left, nobody wins everything, and the honest read is that the best operators run two or three models depending on the asset.

Seedance 2.5 leads on duration and reference density. Best fit: the full 30-second ad, product films, anything where subject consistency across a whole spot is the constraint.
Veo is strongest for audio-native cinematic work and scene-level sound design.
Kling holds the value position and a strong 4K look, and plenty of working creators keep it in the pipeline on cost alone.

Pick per asset, not per subscription. The moment you standardize on one model for everything, you inherit its specific weakness on every brief. Sora's shutdown is the sharper version of that lesson: standardizing on one vendor means their business decision becomes your migration.

What this costs, honestly

ByteDance hasn't published final API rates. Third-party estimates circulating this week land around a fraction of a dollar per second at lower settings, with a full 30 seconds at 4K running considerably more. Treat every number you read right now as an estimate, including that one.

The planning point stands regardless: a 30-second 4K generation isn't something you reroll casually. That changes your workflow. You want your concept, references, and script locked before you spend a full-length generation. Sketch cheap, commit expensive. The setups that let you draft at low resolution and finish at high resolution are the ones that fit a real marketing budget.

How to prepare before it reaches your stack

Seedance 2.5 is rolling out through ByteDance's own apps first, and broad API availability is still landing. That gap is useful. Spend it on the parts that transfer to any model:

Build the reference library. Product angles, creator frames, brand film, voice sample. This is the input that makes 50-reference generation worth anything.
Write your 30-second structure. Setup, tension, product, payoff. If your script only works as four disconnected five-second clips, a one-take model won't save it.
Storyboard before you generate. Cheap frames first, expensive motion second. It's the single biggest cost control in AI video and it works on every model.
Define what "on-brand" means concretely. Not adjectives. Named grade, named lens language, named wardrobe, named voice.

None of that is Seedance-specific, which is the point. The models will keep leapfrogging each other. The teams that win are the ones whose inputs are ready when a new one lands.

Where MITPO fits

We're tracking Seedance 2.5, and it's already visible in the Creative Studio model picker marked as coming soon. We don't wire a model until we can verify it live on a provider we can actually bill through, and no provider carries it for us yet. When that changes, it turns on.

Meanwhile the workflow above is what MITPO is built around: sketch your storyboard first, lock the beats cheaply, then spend real generation budget only on the frames that survived. You can try the live demo without signing up.

The model announcements will keep coming. The prep work is what compounds.