Your first $10 on Sume: model, resolution and length tests

The first $10 on Sume is best spent on three small experiments, not on one big render: a model comparison ($3.48), a resolution comparison ($3.39) and a length comparison ($1.00), then captions on the winner at $0.20. The plan costs $8.06, which leaves $1.94 for retries.

The goal is decisions you can reuse: which model for which job, whether 720p is worth it, and how much a second of length costs. Every figure is a list-rate estimate from the pricing package.

The three experiments

Use one prompt and one start image throughout, so the only variable is the one you are testing. Grok Imagine video on Sume needs an input image, so its line includes a low-quality still. Seedance 2.5 is priced by token count for the resolution and duration; at 720p and 5 seconds it is $2.889, and at 4 seconds it is $2.3112. H3 Max at 768p is priced per second at $0.10, so 5 seconds is $0.50 and 10 seconds is $1.00.

What each test tells you

Test 1 shows the price-versus-look trade between a $0.09 clip and a $2.89 clip, and it is the test most likely to save you money later. Test 2 shows whether 480p is acceptable for your channel; a 480p Seedance clip is 46% of the price of 720p at the same length. Test 3 shows whether longer H3 Max clips look better or just cost more, since price rises linearly with seconds.

Keep the output, write down what you saw, and decide the default model before you spend more.

• Do not test 1080p in the first $10: one 4-second Seedance 2.5 clip at 1080p is $5.69.
• Run the cheap clips first and the expensive one last, so a mistake in the prompt costs little.
• Write the request body once and change one field per test.

Why the reserve matters

Each submit reserves its estimate from the balance, and a submit that cannot be reserved is refused with 402 insufficient_credits before provider work starts, as the generation admission page explains. With $10 you can hold several cheap jobs at once but only a few Seedance jobs, so submit the expensive test last and check the balance with GET /v1/balance first.

After the $10

When the tests are done, write the result as a one-line rule for each decision: default model, default resolution, default length. For example, a rule might read: Grok for b-roll, H3 Max for product shots, Seedance 2.5 720p only for the hero. Those rules become your price sheet. The per-deliverable rate sheet shows what that looks like at scale.

Top up in the dashboard only after you have a rule. A second $10 spent without a decision is the same experiment run twice.

• Record the cost shown on each job next to the result.
• Compare each clip against the cheaper one first; many tests end at the cheapest acceptable model.
• If a test fails on a prompt error, repeat it only once before changing the prompt.

Check your own numbers

The figures here come from the pricing package on the date read. Your requests may differ: an aspect ratio, a reference input or an audio setting can change the estimate. Submit once and compare the job's quoted cost to the table before you scale up.