Seedance Mini, Fast, 2.0 and 2.5: one 5 s clip, four prices

Sume lists four Seedance ids on /v1/videos: seedance-2-mini, seedance-2-fast, seedance-2 and seedance-2.5. They share a request shape and a pricing formula, video tokens, so the choice comes down to what each id costs and what each allows. The surprise is the spread. For one 5-second, 720p, 9:16 clip, the dearest id costs about three times the cheapest.

The prices below come from Sume's pricing tests. Each is the provider list rate times 1.25, applied to 108,000 video tokens, which is 720 x 1280 x 5 x 24 / 1024.

Four prices for one clip

Rates per 1,000 video tokens at Sume are $0.00875 for Mini, $0.014 for Fast, $0.0175 for 2.0 and $0.02675 for 2.5. At 480p the tokens drop to 50,220 and the ladder scales down with them.

What the extra money buys

Seedance 2.5 is the only id in the family that goes past 15 seconds: it takes 4 to 30 seconds. It also runs at 1080p with a higher token rate, so it is the id for long clips and for final renders. Its model card lists eight aspect ratio values including auto, text-to-video, start and end frame image-to-video, and references with image, video and audio.

The 2.0 family is capped at 15 seconds. Mini is the cheapest and is described as cost-biased with fast latency, Fast is the speed-biased row, and plain 2.0 is the quality row at the same envelope. All three accept text, image and reference input and have an optional audio flag.

A sensible workflow

Draft on Mini at 480p, where a 5-second clip costs about 44 cents, then render the keepers on 2.5. Be careful that the same prompt does not produce the same picture on a different model. The draft tells you whether the idea works, not whether the final will match, so keep drafts for composition and wording and not for sign-off.

If your final needs 20 seconds, none of the 2.0 ids can do it. The API answers a 16-second request on seedance-2 with a 400, because it is not in the model's advertised durations, and no job is created.

Check the live catalog

Catalog entries carry the pricing_skus rate per 1,000 tokens as a decimal string, and the completed job carries usage.cost. Use those two numbers rather than this table when you reconcile spend, because price tables move when the provider changes its list.