$50 AI video budget: usable 8-second clips by takes per keeper

$50 buys 50 clips of 8 s on Wan 3.0 at one take, 16 at three; Seedance 2.5 buys 10 or 3. The take rate matters as much as the model. Table inside.

5 min readSume
All posts

A $50 budget buys 50 usable 8-second clips on Wan 3.0 at 720p if every take is kept, but only 16 if each keeper needs three takes. On Seedance 2.5 at 720p the same $50 buys 10 clips at one take and 3 at three. The take rate changes the answer as much as the model does.

Each take is billed when the job is accepted and completes, so an unusable clip costs the same as a keeper. The figures use Sume billable rates (provider list times 1.25), read from the pricing package on 2026-10-03, and an 8-second clip, which is also the default length for Auto create controls in the Video Router docs.

The table

Clip counts are rounded down to whole clips: $50 divided by (price of one take times takes per keeper). One take is the optimistic case; two takes means every second generation is a keeper; three takes means one in three attempts is kept.

Take rates depend on the brief, not just the model. A product on a plain desk with a short prompt may be a one-take job, while a person walking through a crowded scene may need many.

Usable 8-second clips a $50 budget buys, by takes per usable clip (read 2026-10-03)
ModelOne 8 s take1 take1.5 takes2 takes3 takes
Wan 3.0 720p$1.0050332516
Gemini Omni Flash 1.1 720p$1.0050332516
Seedance 2.0 Mini 720p$1.5133221611
Seedance 2.5 720p$4.6210753

Using the table to set a rule

Write your take allowance into the budget before you start. If the first five clips take 2.4 attempts each, the plan should use 2.4 and not 1. Reserve at submit means the money leaves the balance when you press go, not when you pick a winner.

Cheaper models change the right strategy. At Wan 3.0 prices, three takes of an 8-second clip cost $3.00, which is less than one Seedance 2.5 take at $4.62. If the cheap model reaches your quality bar in three takes, it is the cheaper path; if it never does, the expensive one is.

A mixed rule that works

Use a cheap model to explore the shot and an expensive one to finish it. For example, spend two Wan takes ($2.00) to settle the framing and then one Seedance 2.5 take ($4.62) for the keeper: $6.62 per keeper, against $13.87 if you had needed three Seedance 2.5 takes.

The arithmetic is simple, but the habit matters: the number of takes per keeper is the single most useful thing to log from a first week of generation.

What Sume does and does not do

Sume reserves the billable amount on submit and settles on completion; it does not predict how many takes you will need. Confirm rates in the catalog, and read each model's duration limits there.

  • Auto create controls default to 720p and 8 s, 3 to 10 s at 16:9 or 9:16.
  • Seedance 2.5 takes 4 to 30 s per clip; Omni takes 3 to 10 s.
  • Keep a log of takes per keeper.

Rerun the arithmetic

The same line items as a short script, so you can change a quantity and see the total. It is plain arithmetic on the rates in this post; it does not call the Sume API.

# Sume billable USD rates, read 2026-10-03. Edit the quantities.
lines = [
    ('Wan 3.0 720p, 8 s takes (3 per keeper, 16 keepers)', 384, 0.125),
    ('Remaining takes at 1 per keeper, to taste', 0, 0.125),
]
total = 0.0
for label, qty, unit in lines:
    cost = qty * unit
    total += cost
    print(f"{label:<46}{qty:>8g} x {unit:<10g} = ${cost:,.2f}")
print(f"{'Total':<46}{'':>8}   {'':<10}   ${total:,.2f}")

Sources

Related posts

More in Pricing

All Pricing posts

Written by Sume