Seedance 2 at $0.014 per 1,000 tokens: read pricing_skus first

Some video models are priced per second and some per token, and mixing them in one budget is how a bulk run goes over. fal's pricing page, read 2026-10-05, lists H3 Max at $0.05 per second, Kling Video v3 at $0.14 per second and Seedance 2 at $0.014 per 1,000 tokens. Sume's GET /v1/videos/models returns pricing_skus for each model as decimal-USD strings, in the same format OpenRouter uses, so you can read the unit from the catalog instead of from a blog table.

Per second versus per token

A per-second row is easy: seconds times rate. A token row needs a token count, which depends on resolution and duration, so the cost of two clips with the same length can differ. fal's page notes that settings such as resolution, duration and quality may affect the final cost. For a budget, use the catalog's SKU name (for example per-1000-video-tokens) as the unit, and estimate tokens per clip from one real job's usage rather than from a formula you remember.

Read the SKU from the catalog

Sume's catalog values are Sume billable rates, which are provider list times 1.25 for the Video Router rows. This script prints each model id with its SKU map. It needs a key in SUME_API_KEY.

import json, os, urllib.request

req = urllib.request.Request(
    "https://api.sume.com/v1/videos/models",
    headers={"Authorization": "Bearer " + os.environ["SUME_API_KEY"]},
)
with urllib.request.urlopen(req) as resp:
    models = json.load(resp)["data"]

for m in models:
    print(m["id"], m.get("pricing_skus"))

A budget method that survives token pricing

Run one clip, read its usage.cost from the poll response, and divide by the seconds to get an effective per-second figure for that exact recipe. Multiply by the batch and add headroom. As an illustration only, a clip that consumed 100,000 tokens at $0.014 per 1,000 tokens would be $1.40 before any margin. The point is the method: take the unit from the catalog, take the quantity from a real job, and let the per-item generation_spend_cap_usd on a Format bulk run catch the outliers.

• Do not compare a per-token model to a per-second model on the headline rate alone.
• Re-read the catalog on each batch day; the notes in docs carry the date they were read.
• Record usage.cost for every job so next month's estimate starts from your data.