Sora API ended Sept 24: which Sume video model to point code at
The Decoder reported the Sora API ends 2026-09-24. If your code called it, Sume's POST /v1/videos takes a model id; here is how to pick one and re-test.

If your code called the Sora API, it has nothing to call after 2026-09-24. The Decoder reported that OpenAI closes the Sora app on 2026-04-26 and discontinues the API on 2026-09-24, and that users must download their content. The article was written earlier in 2026, so it is a report of a plan, not a status page; check your own account if a call still returns.
Sume does not run Sora and has no Sora model id. What it offers instead is a video endpoint, POST /v1/videos, shaped like OpenRouter's video API, where the model field is a bare Sume id such as kling-3, seedance-2.5, wan-3.0, minimax-h3 or gemini-omni-flash-1.1. Moving off Sora is mostly a prompt and limits re-test, not a rewrite.
What changes in the request when I leave Sora?
The shape you send is a JSON body with a model id and a prompt, and you get a job back that moves through pending, in_progress, completed, failed or cancelled. You poll the job or take a webhook, then download the file. Sume's docs say sync waits are bounded at 30 seconds, so do not design around holding a connection open.
The parts that do change are the limits. Each Sume model has its own duration range, resolutions and aspect ratios, and the catalog example in the docs reports seed as false, so check seed per model before expecting a prompt to reproduce a clip exactly.
| Model id | Duration | Resolutions | Audio |
|---|---|---|---|
| kling-3 | 4-15 s | 720p, 1080p | On or off |
| seedance-2.5 | 4-30 s | 480p, 720p, 1080p | See the model page |
| wan-3.0 | 2-30 s | 480p, 720p, 1080p | See the model page |
| gemini-omni-flash-1.1 | 3-10 s | 360p, 720p, 1080p, 4K | Always on |
| minimax-h3 | 5-15 s | 480p, 768p | See the model page |
Should I pin one model or use sume/auto?
Pin a model id when a client signed off on the look. Use sume/auto when you want Sume to choose; the docs say it does not disclose which family served the request, so it is a poor fit if you must document the model. After Sora, pinning is the safer default: a pinned id changes only when you change it.
What should I re-test first?
Run your five most important prompts on two candidate models and score them yourself. Check the clip length you actually need against the duration column, because a 30-second ask only fits a model whose range reaches 30 s. Then check idempotency: send an Idempotency-Key on every create so a retry does not create a second paid job.
Finally, download finished files to your own storage. The Sora story is a reminder that a hosted URL is not an archive.
Sources
Related posts
More in Developers
- Sora Batch API render queue gone: Sume async jobs instead
OpenAI's Sora guide listed Batch API support before the September 24 shutdown. On Sume the pattern is async jobs, a queue and signed webhooks.
- Sora shutdown lesson: run a one-hour video vendor exit drill
OpenAI discontinued the Sora API on 2026-09-24. Rehearse losing a video vendor: swap the model id, rerun five prompts, and record what broke.
- soundtrack_shorter_than_spine: loop the music bed in Timeline
Timeline warns soundtrack_shorter_than_spine when the music bed is shorter than the audio. Set soundtrack.loop to true; gain, fade and duck defaults inside.
- SQS FIFO 5-minute dedup vs a Sume Idempotency-Key: which wins?
SQS FIFO dedupes for only 5 minutes. A Sume Idempotency-Key covers the create call itself, so use both when a worker can retry after the window.
Written by Sume