Seedance 2.5 price per second: read it from the Sume catalog before Q4

The reliable way to price Seedance 2.5 on Sume is to read its row from GET /v1/videos/models and then bill one 4-second draft to learn your real cost per second. The model accepts 4 to 30 seconds at 480p, 720p and 1080p, and Sume bills the provider list times 1.25.

This post does not quote a per-second figure for Seedance 2.5, because the references it reads do not record one, and a guessed number in a budget is worse than none. The other posts in this series use rates that are written down; this one shows how to get a rate that is not.

What is documented

Sume's Video Router docs list seedance-2.5 with resolutions of 480p, 720p and 1080p and durations of 4 to 30 seconds, with no bitrate_mode. The older Seedance 2.0 family (seedance-2, seedance-2-fast, seedance-2-mini) stays routable by explicit id, but no default or routing preset selects the 2.0 and mini models now. The video docs say Seedance 2.x models accept audio and video references, which most other rows do not.

The catalog response for a model carries pricing_skus, a map from a SKU name to a price string. The docs' own example for Seedance 2.0 shows a key named per-1000-video-tokens, which tells you the unit is token-based, not per-second.

Method: catalog row, then one draft

Step one is the catalog row. Step two is a 4-second 480p draft, which is the cheapest legal request. A completed job returns usage.cost, the Sume billable amount, so divide by 4 for a draft rate. Repeat once at 1080p to see how resolution scales, and keep both numbers in your budget sheet.

import os, requests

BASE = "https://api.sume.com/v1"
H = {"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"}

models = requests.get(f"{BASE}/videos/models", headers=H).json()["data"]
row = next(m for m in models if m["id"] == "seedance-2.5")
print(row["supported_resolutions"], row["supported_durations"][:2], row["pricing_skus"])

job = requests.post(
    f"{BASE}/videos",
    headers={**H, "Idempotency-Key": "seedance-price-probe-480p"},
    json={"model": "seedance-2.5", "prompt": "A gift box opening on a snowy table",
          "resolution": "480p", "duration": 4},
).json()
print("poll:", job["polling_url"], "then read usage.cost and divide by 4")

Turn the probe into a budget

Once you have a draft rate, build the sheet the same way as the rows in the other posts: rate × seconds, rounded up per job, then clips × price. Suppose your 4-second 480p probe returns a cost of X dollars. Then ten 4-second drafts cost 10X, and a 15-second 1080p final costs 15 × (your 1080p rate). The formula is simple, and the number is the only unknown.

Why this beats a published number

Provider prices move, and the Sume docs name the date of each recorded list rate for a reason. A rate you measured today, from the same API your production job will call, includes the 1.25 factor and the rounding, which a number copied from a blog post does not. Idempotency keys make the probe safe to repeat.

If the figure matters to a client quote, ask the admission preview route what a request would reserve before you submit: the lip-sync docs name POST /v1/generation/admission-preview for exactly that purpose, and it takes the public id.