Seedance 2.5 Generates 30 Seconds in One Pass
ByteDance doubled its clip length to 30 seconds and takes 30 images as reference. What Seedance 2.5 does, what it costs, and when it's worth the price.
By Yuvraj Singh·Founder, Leaxor
Sources linked throughoutClaims checked against vendor docsReviewed August 2026
Short version: ByteDance released Seedance 2.5 on 31 July 2026. It generates 30 seconds in a single pass, double what Seedance 2.0 managed, and accepts an unusual amount of reference material in one request. It is also the most expensive model on this shelf, and for most work that is not a good trade.
Disclosure: Leaxor is mine and runs it.
What ByteDance shipped
TechNode reported the release on 31 July with three headline capabilities: a 30-second clip in a single run, multi-turn extension for longer sequences, and reference input of up to 30 images, 10 videos and 10 audio clips in one request.
The lineage matters for context. Seedance 2.0 topped the Artificial Analysis video arena for both text-to-video and image-to-video in early 2026. This is not a lab catching up.
Thirty seconds is the headline, but reference input is the real change
Length gets the attention. Doubling from 15 to 30 seconds in one pass removes the stitching problem for a whole class of work, because a single continuous generation has no seams to hide.
The reference capacity is the more interesting number though. Thirty images, ten videos and ten audio clips in one request is an order of magnitude more than anything comparable accepts.
That matters because consistency is the wall everyone hits. Getting the same character to look like the same character across eight separate clips is the single most common reason people abandon AI video. More reference input is a direct attack on that problem rather than a spec-sheet number.
Whether it solves it is a different question, and one nobody has enough public output to answer yet. Treat the claim as promising rather than proven.
Where you can actually use it
Availability was staged. It rolled out first inside ByteDance's own products, Jimeng AI and the Pro tier of Doubao, with API access following on Volcano Engine's Ark platform and third-party hosts.
That pattern is becoming the norm. Ship into your own apps first, open the API second. Google did the same with Nano Banana, and Meta did it with Muse.
What it costs, and why the shape is odd
Seedance 2.5 is the priciest model here. Its rates are published on fal's catalogue alongside everything else.
| Duration | Credits | Cost |
|---|---|---|
| 5 seconds | 23 | $2.30 |
| 8 seconds | 36 | $3.60 |
| 10 seconds | 45 | $4.50 |
| 20 seconds | 89 | $8.90 |
Pricing is linear per second, so a thirty-second continuous run lands near 133 credits, or $13.30. That is a lot for one clip and almost nothing else can produce it at all.
Which is the honest way to use this model: reach for it when the length is the point. Using it for five-second clips means paying four times the Wan 3.0 price for something Wan does adequately.
What the reference input actually changes
Thirty images, ten videos and ten audio clips in one request is the number people quote. What it is for gets less attention.
Every AI video workflow hits the same wall: the character in shot four does not look like the character in shot one. The usual workarounds are all bad. You regenerate until you get a near-match and accept the drift. You describe the character in exhaustive prose and hope. You give up on multi-shot sequences entirely and make single clips.
Feeding thirty reference images attacks that directly. Instead of describing a face, you show it thirty times from different angles and let the model work out what is invariant.
Whether that holds up across a full sequence is the open question, and there is not enough public output yet to answer it. Treat it as the most promising attempt on the hardest problem in the field, not as a solved one.
Multi-turn extension, and why it is the quieter feature
Beyond the thirty-second single run, the model supports continuing an existing clip rather than starting a new one.
That is a different operation from generating a fresh clip and cutting it on. An extension inherits the state of the frame it continues from, so the lighting, the subject and the camera position carry over instead of being approximated again from a prompt.
For anyone who has tried to stitch two separately generated clips and watched the subject change coats at the join, this is the more useful half of the release.
Where it sits
| Model | Max single clip | 5 seconds |
|---|---|---|
| Seedance 2.5 | 30s | 23 cr ($2.30) |
| Wan 3.0 | 30s | 5 cr ($0.50) |
| Flux 3 | 20s | 17 cr ($1.70) |
| MiniMax H3 Max | 15s | 5 cr ($0.50) |
| Gemini Omni Flash | 10s | 3 cr ($0.30) |
Wan 3.0 also reaches 30 seconds at a fraction of the price, which is the comparison that should make you think. The case for Seedance is quality and reference handling, not duration alone.
Should you use it?
Yes for one continuous shot beyond fifteen seconds where quality carries the piece.
Yes if you have real reference material and need the output to respect it.
No for short clips. You are paying four times the price for something the cheap tier does adequately.
No for exploration, obviously. Draft on Gemini Omni Flash at $0.30 and come here once.
Seedance 2.5 is live in Leaxor's clip generator, credits at $0.10 with no subscription required and no expiry. The full shelf and where each model earns its price is in the per-second cost breakdown, and the August roundup covers everything else that shipped this month.
Reading this with an AI assistant? Ask it to summarize this in ChatGPT, Claude, or Perplexity.
Frequently asked questions
What is Seedance 2.5?+
Seedance 2.5 is ByteDance's video generation model, released on 31 July 2026. It generates a 30-second clip in a single run, double the 15-second ceiling of Seedance 2.0, and supports multi-turn extension for longer sequences.
How long can Seedance 2.5 videos be?+
Thirty seconds in one pass, which is the longest single-run generation among the models covered here. Beyond that it supports multi-turn extension, where you continue an existing clip rather than starting again.
How much does Seedance 2.5 cost?+
It is the most expensive model on this shelf, though not by the margin it once was. On Leaxor a five-second clip is 23 credits, which is $2.30, ten seconds is 45 credits and a full thirty-second run is around 133 credits. Credits are $0.10 each with no subscription required.
How many reference images can Seedance 2.5 take?+
Reporting at launch described up to 30 images, 10 videos and 10 audio clips as reference material in a single input. No other model currently accepts anything close to that volume of reference, which is the feature that most distinguishes it.
Is Seedance 2.5 better than Seedance 2.0?+
On capability, clearly: double the length, far more reference input, and multi-turn extension. Seedance 2.0 topped the Artificial Analysis video leaderboards in early 2026 for both text-to-video and image-to-video, so 2.5 builds on a strong position rather than fixing a weak one.
When should I use Seedance 2.5 instead of a cheaper model?+
When you need one continuous shot longer than fifteen seconds, or when you have a lot of reference material and need the output to respect it. For everything else the cheaper models produce comparable results at a fifth of the price.
Try the models yourself
Leaxor runs the leading image and video models in one place, paid per credit.
Get startedYuvraj SinghFounder, Leaxor
Built Leaxor, an all-in-one AI image and video generator, to kill the two bottlenecks in publishing: 3–5 hours per short, and a second tool for every thumbnail. Now both take minutes.