Fan out four Omni Flash clips in parallel and poll the jobs

To make a long piece from short Omni clips, submit one request per beat at once, store each job id, and poll each job's status until it is terminal. OpenAI's Sora guide says stitching shorter clips follows instructions more reliably than one long generation, and Sume's job model suits that: every generation returns a durable job.

Sume details are from the Jobs and results docs and Video Router docs; the Sora point is from OpenAI's Sora 2 Prompting Guide (read 2026-10-03).

Submit each beat with its own key

Writes need an Idempotency-Key. Use one per beat, such as story-001-beat-1, so a retry of the same beat is safe while a different beat is a different request. Admission limits can leave a job queued; the docs call that a normal accepted state.

Poll with backoff

Read the job id from each submit response and keep it. The docs say to use exponential backoff, stop on a terminal status, and not resubmit a paid request just because a local process timed out.

curl https://api.sume.com/v1/jobs/$JOB_ID/status \
  -H "Authorization: Bearer $SUME_API_KEY"

Fetch results and join

GET /v1/jobs/:id/result works only after completion; before that it returns a conflict. When all four are complete, pass the clip URLs to a Timeline 1.0 render. If you prefer push to poll, Sume also sends job.completed webhooks, signed with HMAC over the timestamp and raw body.

Cost shape

Four 8 second 720p clips are 32 seconds of output: 32 x $0.125 = $4.00 on Sume at list x 1.25, plus $0.10 for a one minute timeline render. See the Video Router docs.