Seedance 2.5 Flash Video Model Opens on APIs

ByteDance has opened Seedance 2.5 flash through its Volcano Ark and BytePlus ModelArk APIs. Volcano Engine’s model-release announcement lists doubao-seedance-2-5-flash-261115 as a new video-generation model under November 2026. The entry says the model supports text-to-video, first-frame and first-last-frame generation, omni-reference video, video editing and video extension, at 480p and 720p, and that it has no draft mode. BytePlus ModelArk’s Seedance 2.5 series tutorial, last updated on 15 November 2026, lists the same model for BytePlus developers as dreamina-seedance-2-5-flash-261115. The Create video generation task API page now says Dreamina Seedance 2.5 and Seedance 2.5 flash “are now fully available.”

The opening is aimed at developers who want a 2.5-series video model on those APIs without paying the flagship’s token rate or using its 1080p and draft-mode options. It follows ByteDance’s release of Seedream 5.5, the company’s latest image-creation model, on the same platforms.

Flash generates 4 to 30-second clips at 480p or 720p

The BytePlus tutorial says the Seedance 2.5 series can “generate up to 30 seconds of video in one request.” That is twice the 15-second cap on Seedance 2.0. Output duration for flash is 4 to 30 seconds, or -1 for a duration the model chooses inside that range. Resolution is 480p or 720p, both at 8-bit colour depth. Seedance 2.5 itself adds 1080p at 10-bit. Flash also has no draft mode; the capability table marks draft as available only on Seedance 2.5.

The supported task types match the Volcano announcement: text-to-video, first-frame and first-last-frame generation, omni-reference, video editing and video extension. The model can also generate video with audio. Aspect ratios are listed as 21:9, 16:9, 4:3, 1:1, 3:4, 9:16 and adaptive, and output formats are mp4 and mov. Offline inference is not supported.

Reference limits are tighter on flash than on the flagship. A flash request can include up to 10 reference images and up to 5 reference audio or video assets, for a total of 20. Seedance 2.5 allows 30 images and 10 audio or video assets. Both 2.5-series models raise those caps from Seedance 2.0, which allowed 9 images and 3 audio or video assets, and they raise the maximum total duration of reference audio or video from 15 seconds to 30 seconds. The Chinese-language Volcano Ark tutorial carries the same limits.

Before a Seedance 2.5 series model can be enabled, BytePlus says a developer should have a BytePlus account balance above USD 30, a dedicated AI Savings Plan at the USD 30 tier or above, or a Dreamina Seedance 2.5 series resource pack with available quota. The Volcano Ark tutorial sets the matching China-side threshold at an account balance above 200 yuan, a 200-yuan-tier savings plan, or a Seedance 2.5 series resource pack.

The 30-second request window is now a common published ceiling among recent video models. HeyGen’s MiniMax H3-based video model, which TSN covered when that product moved to 30-second clips, sits in the same duration band, though it is a different company’s service and is not a ByteDance API.

The published $7.0 and $4.3 figures are token rates

BytePlus ModelArk’s pricing page puts Dreamina Seedance 2.5 flash, for online inference at 480p and 720p, at $7.0 when the input has no video and $4.3 when the input includes video. Those figures use the page’s general unit: “Token prices are USD per million tokens.” The video-billing rule is “Estimated video price: Token unit price × token consumption.” They are not a price per video.

The same page gives worked examples for a 5-second, 16:9 clip with no video input: $0.34 per video at 480p, or $0.07 per second, and $0.76 per video at 720p, or $0.15 per second. When the input includes video, the listed 5-second 16:9 ranges are $0.37 to $1.45 at 480p and $0.84 to $3.25 at 720p, depending on input duration between 2 and 30 seconds. A note on the page says that, under otherwise identical conditions, “a video generated with video input costs more than one generated without video input,” even though the with-video token rate is lower, because token consumption also counts input duration.

For comparison on the same table, flagship Seedance 2.5 is listed at $10.70 without video input and $6.40 with video input at 480p and 720p, and it is the only 2.5-series model priced at 1080p. Seedance 2.0, at 480p and 720p, is listed at the same $7.0 and $4.3 token rates as flash. Failed generations, the page says, are not charged when the failure is due to reasons such as content moderation.

Developers who want a video model through a third-party API rather than ByteDance’s own platforms have a separate route in fal’s FLUX 3 Max video model, which TSN covered when fal launched it with Black Forest Labs. That is a different product, on a different platform, with its own price list.

The $7.0 figure is a token rate, not a price per clip

The easiest misreading of the BytePlus table is to treat $7.0 or $4.3 as the cost of a clip. The page’s own unit and billing formula point the other way. A clip is priced as a token rate times the tokens it consumes, and those tokens scale with resolution, duration and whether a video is included in the input. The $0.34 and $0.76 figures are the company’s own examples for one narrow case: a 5-second 16:9 output with no video input, at 480p and 720p. They are a useful yardstick, and they are not a flat fee for every request.

The lower with-video token rate can also mislead if it is read in isolation. Officially, a request that includes video still tends to cost more than one that does not, because the estimated token formula adds input duration to output duration. The worked ranges for a 5-second 16:9 output with video input start slightly above the no-video examples and rise as the input gets longer, up to $1.45 at 480p and $3.25 at 720p for a 30-second input. Anyone costing an editing or extension workflow has to use those input-inclusive figures, not the $0.34 and $0.76 no-video examples.

Flash is also a narrower product than its name might suggest beside Seedance 2.5. It shares the 30-second request window and the same task list, including editing and extension, but it stops at 720p, it has no draft mode, and it accepts fewer reference assets. The flagship remains the 1080p and draft-mode option, at a higher token rate. The official pages also publish no generation-speed claim for flash.

The activation thresholds are part of the same practical picture. A BytePlus balance above USD 30, a matching savings plan, or a resource pack is required before the 2.5 series can be enabled; Volcano Ark’s equivalent is 200 yuan. That is a platform condition, not a per-clip price, and it sits in front of the token meter rather than instead of it.

What is now on the record is narrower, and more useful, than a marketing comparison: two official model IDs, a 4 to 30-second window at 480p or 720p, a published per-million-token rate, and worked examples for a 5-second 16:9 clip. Those are the figures to cost against until BytePlus or Volcano Engine changes the table.