A 75-second AI video in three jobs (30 + 30 + 15): Wan vs Seedance

A 75-second video is three jobs of 30, 30 and 15 seconds: $9.38 on Wan 3.0 at 720p, $43.34 on Seedance 2.5 at 720p, before a small join.

4 min readSume
All posts

A 75-second video is three jobs of 30, 30 and 15 seconds, and it costs $9.38 on Wan 3.0 at 720p and $43.34 on Seedance 2.5 at 720p, before a small join. The longest clip either model makes is 30 seconds (Video generation), so anything longer is a chain.

Cost of the 75 seconds

Sume bills the provider list rate multiplied by 1.25 on every video model (Video Router). The figures here come from the price estimators in Sume's repository, run on 2026-10-07, and are rounded to the cent. The live numbers for your workspace are in the catalog response and on the API pricing page.

75 seconds as 30 + 30 + 15 on Sume (read 2026-10-07)
ResolutionWan 3.0Seedance 2.5
480p$4.69$20.15
720p$9.38$43.34
1080p$18.75$106.62

Why 30 + 30 + 15

Splitting into the fewest jobs keeps the number of joins low and the number of visible seams at two. Each seam is a place where the look can shift, so give the next segment the last frame of the previous one as its first frame. Both models take a first-frame image.

ByteDance's own description of Seedance 2.5 is 30-second single-pass audio-video clips (Seed blog, read 2026-10-07), which is why the chain is built from 30-second pieces.

  • Pay for the last segment at 15 seconds, not 30: you save $1.88 on Wan 720p.
  • Generate segments in order if continuity matters; in parallel if it does not.
  • Keep resolution identical across the three.

Reserve for the chain

Sume reserves the estimated amount when it accepts a job, captures it on success, and refunds the reservation if the job fails or is cancelled before capture (Usage, Generation admission). So the balance you need on hand at submit time is the price of everything in flight, not only the final total.

A Pro workspace runs 4 jobs at once, Startup 8 and Scale 20 (the plans page lists the ladder), so a big batch drains in waves rather than all at once.

Set a cap before you start: from MCP, a paid tool takes dry_run=true and max_spend_usd (MCP tools and gates). Send a distinct Idempotency-Key per job so a retry replays the original job instead of charging again.

Sources

Related posts

More in Pricing

All Pricing posts

Written by Sume