Seedance 2.5, 2.0, Fast and Mini: 5 s at 720p on Sume, $0.95 to $2.89

Sume lists four Seedance rows. For a 5 s clip at 720p, Seedance 2.5 costs $2.89, Seedance 2.0 costs $1.89, Seedance 2.0 Fast costs $1.52 and Seedance 2.0 Mini costs $0.95. Only 2.5 reaches 30 seconds; the others stop at 15 s, and Fast and Mini stop at 720p.

Same tokens, different rates

All four rows count tokens the same way (width x height x seconds x 24 / 1024), so a 5 s 720p clip is 108,000 tokens on each. The list rate per 1,000 tokens in SEEDANCE_VIDEO_USD_MICROS_PER_1000_TOKENS is what differs: 21,400 for 2.5, 14,000 for 2.0, 11,200 for Fast and 7,000 for Mini. Sume then multiplies by 1.25 and rounds up to the cent.

What you give up

The Video Router docs give the envelope: seedance-2.5 runs 480p, 720p and 1080p for 4 to 30 s. seedance-2 runs the same resolutions for 4 to 15 s. seedance-2-fast and seedance-2-mini run 480p and 720p for 4 to 15 s. The 2.0 family keeps reference image, video and audio inputs, so a Mini test is a faithful dry run of a reference set.

The docs state that no default or routing preset selects seedance-2 or seedance-2-mini now. They stay routable when you request the id explicitly.

A test-then-finish routine

Use Mini to check that references load and that the prompt composes. Move to 2.5 for the final if you need more than 15 s or 1080p. Because the API has no seed, the final is a new take, so judge motion and acting on the final itself, not on the Mini draft.

For a 100-clip batch the gap matters. Ten test clips of 5 s at 720p cost $9.50 on Mini and $28.90 on 2.5.

Listing the rows yourself

Read the live catalog before you budget, since list rates can change.

curl -s https://api.sume.com/v1/videos/models \
  -H "Authorization: Bearer $SUME_API_KEY" \
  | jq -r '.data[] | select(.id | startswith("seedance")) | [.id, (.supported_durations | "\(min)-\(max) s"), (.supported_resolutions | join("/"))] | @tsv'

Choosing a row

Pick 2.5 for length (up to 30 s) and 1080p, 2.0 for 15 s clips at 1080p with a lower rate, Fast for iteration at 720p, Mini for prompt and reference tests. Every row takes the same request body, so moving between them is a one-field change. If you route from code, keep the model id in config and log usage.cost for each job so a rate change shows up in your data.

Because each of the four rows is explicit, none of them is what sume/auto serves. Auto resolves to Gemini Omni Flash 1.1 for ordinary generation.