100-clip batch on a Pro workspace: 24 accepted jobs, 18-job waves

A Pro workspace accepts at most 24 paid generation jobs at once (4 processing plus 20 queued), so a 100-clip batch cannot be submitted in one burst. Using the 18-job wave_size_hint from an empty queue, it takes 6 waves (5 x 18 = 90, plus 10). The clips themselves cost 100 x 6 s x $0.125 = $75.00 on Wan 3.0 at 720p, whichever way you pace them.

The limits, from the admission doc

The numbers below are the documented defaults. Your workspace's generation_limits field in a submit response is the source of truth, because admin overrides can raise the cap.

Waves and dollars

With the hint of 18, 100 clips make 6 waves: 18, 18, 18, 18, 18, 10. A full first wave reserves 18 x $0.75 = $13.50 (a six-second Wan 720p clip is 6 x 0.125 = $0.75). If you fill all 24 slots at once, the reservation is 24 x $0.75 = $18.00.

The docs warn that the hint is a submission-wave hint, not a concurrency limit. If you keep the in-flight work inside the processing cap, the budget is 4, and 100 clips take 25 refills of 4. Both are valid; the first fills the queue, the second keeps it short.

Over-submitting is not a billing event: past 24 accepted jobs Sume returns 429 queue_full and any hold on the rejected job is released. Retry it with the same idempotency key.

Cumulative spend by wave

With six-second Wan 720p clips at $0.75 each, waves 1 to 5 each hold $13.50 (18 x 0.75) and wave 6 holds $7.50 (10 x 0.75). Cumulative: $13.50, $27.00, $40.50, $54.00, $67.50, $75.00.

Suppose the balance is $60.00. Four waves use $54.00, leaving $6.00, which covers 8 more clips (6.00 / 0.75), so the batch stops at clip 80 with 402 insufficient_credits. If you had submitted all 100 without pacing, the first 24 would have been accepted and the balance check would have failed later in the same way, with a messier set of half-finished work.

So pace in waves, check the balance between them, and write down which clip indexes completed. A retry with fresh idempotency keys for the missing 20 clips costs 20 x $0.75 = $15.00 once the wallet is topped up.

A guard for the wallet side

Before each wave, read GET /v1/balance and compare it with wave size x clip price. A $75.00 batch against a $60.00 balance fails on the 81st clip with 402 insufficient_credits, so top up first or cut the batch to 80. Fail-closed on both limits: stop on 402, back off on 429.

Checklist before the first wave

Check the numbers in your own workspace rather than the table above.

• Read generation_limits from the first submit response and compute max(0, concurrency_limit - active - queued) as the in-flight budget.
• Store one idempotency key per clip index, such as batch-014-clip-072, so a retry after 429 or 503 repeats the same job.
• Log the cumulative reservation after each wave. For this batch it should read $13.50, $27.00, $40.50, $54.00, $67.50 and $75.00.
• Stop the whole batch on the first 402, not just that clip: the rest will fail the same way until the wallet changes.