After Sora: keep the recipe, not the prompt, as your stable layer

OpenAI named no replacement for the Sora Videos API. Put house style and output rules in a Sume Format so the next model change does not rewrite your code.

4 min readSume
All posts

If the Sora shutdown taught you anything, it is that a prompt tied to one model is the most fragile thing in a video pipeline. The better layer to keep is the recipe: house style, output contract and quality bar, stored once and reused. A Sume Format stores that recipe, so your backend sends a small input and the model underneath can change.

What actually ended

OpenAI's deprecations page lists the Videos API and five Sora model ids as deprecated, with removal on September 24, 2026, announced March 24, 2026. The replacement column names nothing. That leaves each team to choose its own successor.

Sora entries on the OpenAI deprecations page (read 2026-10-07)
ItemStatus
Videos APIDeprecated, removal September 24, 2026
sora-2, sora-2-proDeprecated
sora-2-2025-10-06, sora-2-2025-12-08, sora-2-pro-2025-10-06Deprecated
Recommended replacementNone listed

Where the prompt fails you

A prompt that reads well for one model often needs rewriting for the next: different length limits, different handling of references, different audio behavior. If each call site builds its own prompt string, every call site needs the rewrite. If the prompt lives in a single recipe, one edit fixes all callers.

What a Format holds

A Format is a saved production recipe: a house style, an output contract and a playbook for one kind of video. Your backend calls it by name with POST /v1/formats/{handle}/{slug}/runs. The recipe is set before your instruction, so you do not resend a system prompt on every call. If your instruction and the recipe disagree, the model follows your instruction (Format API).

  • Style and rules live in the Format, versioned. Each receipt shows which format.version ran.
  • Your per-call data lives in input: product URL, script, host image.
  • You can bind a JSON Schema so the result comes back in your own shape (Structured output).
  • A spend cap per run limits the cost of a bad recipe.

When a Format is the wrong layer

A Format run is an agent turn that takes minutes, not a single model call. If you only need one clip from a prompt, POST /v1/videos with a model id or sume/auto is simpler (Generate videos). Use a Format when the job has several steps, such as host takes, B-roll, voiceover and assembly, and when you want the same look every time.

Moving to a Format also does not make output identical to what Sora produced. Expect to re-judge quality on a sample set, and keep the old files as a reference.

A migration order

First move the call sites that share a style into one Format. Second, test it against ten of your old prompts. Third, set a spend cap and a webhook, then switch traffic. The next time a model is retired, you edit the recipe once and read the receipts.

Sources

Related posts

More in Formats

All Formats posts

Written by Sume