Sora API shutdown report: which claims Sume docs can answer

Pondero reports that all calls to OpenAI's Sora models began returning errors on 2026-09-24 and that the deprecations page lists no replacement. It names Runway (Gen-4 Turbo), Kling (v2 Pro) and Seedance as the closest alternatives, and says none is a drop-in. Of those three, the Sume docs list Seedance models. They list no Runway model, and this post makes no claim about Kling v2 Pro.

That split matters for planning. A migration plan that says "use the three named" fails if one of them is not on the platform you chose. Check the catalog first, then write the plan.

Claim by claim

The table sorts what the report says from what the Sume docs let us state. The right column is limited to what the docs show.

What to do with the gap

The report's point that pricing, prompt formats and resolution options differ is the practical one. On Sume, the video request has model, prompt, duration, resolution and aspect_ratio, and the limits differ by model, so a Sora-shaped request needs a field-by-field port, not a base URL swap. The docs also say that size, seed and provider options return a 400 unsupported_parameter on this surface.

Read GET /v1/videos/models and pick from what comes back. Keep a second model in the chain if the first one is critical, and size every request to the overlap of the two. The posts on duration windows and the Node validator show how.

What the report cannot tell you

It is a news item, so it carries no prices for any replacement and no latency figures. Do not read a ranking into the order the models are named. Budget from the catalog's own prices and test your three real prompts on each model before you commit a pipeline.

A short pre-migration checklist

• List every Sora field your client sends today, then mark each as supported, renamed or rejected on the new surface.
• Read the catalog for the model you pick and write its duration, resolution and ratio lists into your config.
• Choose a second model for the same prompt and check that its limits overlap with the first.
• Send three real prompts through both and compare the clips and the usage.cost values.
• Add polling or a webhook, plus an Idempotency-Key on every submit, before you move traffic.

Where the report stops

The report also covers OpenAI's reasons and the app's downloads. None of that changes what a Sume request looks like, and none of it is a Sume fact. Only the shutdown dates and the named alternatives matter for the port, and both are scoped here to the report that stated them.