Re-render 300 old Sora prompts: queue capacity by plan

To re-render a few hundred saved Sora prompts on Sume, pace submissions to your workspace's accepted-job capacity. The generation admission guide lists processing concurrency and queue capacity by plan: for example Pro accepts 24 paid jobs at a time and Scale 120. Past that, submits fail with 429 queue_full. Send a wave, wait for capacity, send the next, and reuse the same Idempotency-Key on any retry.

Know your ceiling

Concurrency is a dispatch limit, not a submit limit: extra valid jobs wait as queued. The queue capacity is the real ceiling for a burst. The numbers below are the documented defaults; the dashboard Concurrency tab and the generation_limits field are the source of truth for your workspace.

Pace by the snapshot, not a constant

Every submit response includes generation_limits when Sume can compute it. The guide gives a rule for new in-flight work: max(0, concurrency_limit - active_generation_jobs - queued_generation_jobs), capped by queue_capacity_remaining. It also says wave_size_hint is only a hint for submission waves and never a concurrency limit. On Pro with nothing running, the headroom rule gives a budget of 4 in-flight jobs, while the hint (floor of 24 times 0.75) is 18. Pick the conservative number when you want to avoid long queues.

A plan for 300 prompts

Number the prompts, derive a stable Idempotency-Key per prompt, and write the job id next to each row as soon as the submit returns. Submit one wave, poll the batch, and top up as jobs reach a terminal state. On queue_full, stop adding work, cancel queued jobs you no longer need, and retry after capacity opens with the same key.

At 24 accepted jobs, 300 prompts means about 13 waves if each wave fully drains before the next. Real throughput is higher, because you refill as jobs finish.

• Persist the key and job id before the next submit.
• Honor retry-after when present.
• Cancel only works before generation starts; later it returns 409 job_generation_already_started.

Do not skip the money check

A wave also needs balance. Submit fails with 402 insufficient_credits when Sume cannot reserve the estimate, so read GET /v1/balance before each wave. The video generation docs cover the request shape; keep prompts and models in a config file so the job is restartable.