How many 480p drafts cost one 1080p render? Wan 3.0 and Seedance 2.5

On Wan 3.0 a 9-second 1080p render costs $2.25 and a 480p draft costs $0.5625, so one 1080p render buys exactly 4 drafts. On Seedance 2.5 (9:16, 9 seconds) the ratio is 5.29: $12.794 for 1080p against $2.4181 for 480p.

That ratio is the whole draft-first decision. If you will review more than the break-even number of takes per shot, draft low and render the keeper high. If you accept the first take most of the time, skip the draft.

The two ratios, priced

Both rows are one 9-second clip. The Seedance 2.5 row is 9:16 because Seedance 2.5 is token-priced and depends on aspect ratio; Wan 3.0 is a flat per-second rate.

Drafts plus one final versus rendering at 1080p every time

Table below compares N takes. The left cost columns are N drafts at 480p plus one 1080p final; the right columns render all N takes at 1080p. Arithmetic: N x $0.5625 + $2.25 for Wan 3.0; N x $2.4181 + $12.794 for Seedance 2.5.

Reading the break-even

On Wan 3.0, four 480p drafts plus a 1080p final cost $4.50, which is the same as two straight 1080p renders ($4.50). So drafting only pays off if you would otherwise render a third 1080p take. For Seedance 2.5 four drafts plus a final cost $22.4663 against $25.5879 for two 1080p renders, so the tie moves later, because the 480p draft is a smaller share of the 1080p price: even five drafts plus a final ($24.8844) stay under two 1080p renders.

Put differently, drafts are a bet that your prompt needs several attempts. Measure your own first-take acceptance for a week before you build a draft stage into the pipeline.

What to check before you trust the draft

Seedance 2.5 and Wan 3.0 do not take a seed in the Sume catalog (the catalog reports seed: false for the v1 models), so the 1080p final is a new take, not an upscale of the draft. A draft tells you whether the prompt and the reference work; it does not promise the same motion at 1080p.

Check each model's supported_resolutions and supported_durations first, because limits differ per model (Wan 3.0 accepts 2 to 30 seconds, Seedance 2.5 accepts 4 to 30).

• One draft cost is wasted whenever the final is also re-rolled, so count the final's own retries separately.
• Send a fresh Idempotency-Key for each new take; reuse a key only for an exact retry of the same request.
• Billing is a reserve of list x 1.25 at submit and a capture on success, so a draft that fails is released, not charged.

curl -X POST https://api.sume.com/v1/videos \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: shot-07-draft-1" \
  -d '{"model":"wan-3.0","prompt":"Slow dolly toward a ceramic mug on a desk","resolution":"480p","duration":9}'