Sora 2 API Shuts Down Sept 24: Migration Guide to Seedance 2.5 & Veo 3.1

OpenAI is switching off the Sora API on September 24, 2026. The consumer app went dark months ago, on April 26 — so if Sora is still in your stack today, it is there as an API dependency, and that dependency has days left.

This is a migration guide, not an obituary. It maps what you were asking Sora for onto the two models that can actually answer: Seedance 2.5 and Veo 3.1. Every capability number below comes from the model registry Imgo routes against, so what you read here is what the API will accept.

What is actually shutting down, and when

Two dates, two different things, and conflating them is why some teams think they already migrated:

Surface	Date	Status
Sora web and app experiences	April 26, 2026	Already discontinued
Sora API	September 24, 2026	Shutting down

Both dates are OpenAI's own, published in their help centre article on the Sora discontinuation. After the API goes, generated content and account data follow: OpenAI has said it will permanently delete data associated with Sora use once any final export window closes. Export anything you still need before you migrate.

Pick your replacement by what you were using Sora for

Sora 2 was a text-to-video model with a deliberately narrow surface: three aspect ratios, three clip lengths, no frame conditioning, no reference images. That narrowness makes the migration decision simple, because almost anything replacing it is a superset.

	Sora 2 (ending)	Seedance 2.5	Veo 3.1
Clip length	4 / 8 / 12s	5 / 10 / 15 / 20 / 25 / 30s	4 / 6 / 8s
Resolution	Fixed	720p / 1080p / 4K	720p / 1080p / 4K
Aspect ratios	16:9, 9:16, 1:1	16:9, 9:16, 1:1, 4:3, 3:4, 21:9	16:9, 9:16
Start / end frame	No	Yes	Yes (end frame requires a start frame)
Reference images	No	Up to 50	Up to 3
Reference video	No	Yes	No
Native audio	No	Yes	Yes
Seed	No	No	Yes

Read that table as two different answers rather than a ranking:

• Go to Seedance 2.5 when length or identity matters — anything past eight seconds, vertical and square deliverables from one source, or a character and product that must survive across shots. Fifty reference slots and reference-video input are the reason.
• Go to Veo 3.1 when the shot is short and has to look expensive, and when you need reproducibility. It is the only one of the three with a seed, so a prompt you approve today renders the same way next quarter.

Two constraints worth knowing before you write the request, because both will otherwise surprise you at runtime:

• Veo renders 8 seconds whenever the output is 1080p or 4K, or whenever the run starts from a frame — whatever duration you asked for. Budget for 8s at those settings.
• Veo accepts an end frame only alongside a start frame. Sending an end frame on its own is a request-shape error, not a fallback.

The migration, concretely

A Sora 2 call carried a prompt, an aspect ratio and a duration. Here is the same intent against Imgo, which routes to either model through one request shape:

POST /api/tools/video/generate
Content-Type: application/json

{
  "model": "seedance-2.5",
  "prompt": "A ceramic mug on a linen cloth, morning light, slow push in",
  "aspectRatio": "9:16",
  "resolution": "1080p",
  "duration": 10,
  "generateAudio": true
}

Switching models is the model field and nothing else:

{
  "model": "veo-3.1",
  "prompt": "A ceramic mug on a linen cloth, morning light, slow push in",
  "aspectRatio": "9:16",
  "resolution": "1080p",
  "duration": 8,
  "seed": 42
}

Three mechanical changes cover most Sora migrations:

1. Durations. Sora's 4 / 8 / 12 does not map cleanly onto either replacement. A 12-second Sora clip becomes 10 or 15 on Seedance 2.5; on Veo the ceiling is 8, so a 12-second shot has to become two.
2. Resolution is now yours to choose. Sora fixed it. Both replacements expose 720p / 1080p / 4K, and the choice moves your cost — so set it deliberately rather than inheriting a default.
3. Audio arrives with the video. Sora returned silent clips and most pipelines bolted audio on afterwards. Both replacements generate a native soundtrack. If you already have a voiceover step, turn generation off with "generateAudio": false rather than paying for audio you discard.

Don't pick one — route between them

The reason teams pinned themselves to Sora was never that Sora was the best model. It was that rewiring a video pipeline is expensive, so whichever model you integrate first tends to become the one you keep. That is exactly the position that made this shutdown painful.

Imgo exists at that seam: one request shape, one credit balance, and a model field you can change per shot. Send the same prompt to Seedance 2.5 and Veo 3.1, compare the two takes side by side, and keep whichever one the shot needed — without writing a second integration or holding a second API key.

Which matters more the week after a shutdown than at any other time. The next deprecation is not going to be announced on your schedule either.

A migration checklist for this week

1. Export first. Anything still living in Sora should be pulled down before the API closes; deletion follows the shutdown.
2. Grep for the model id. Find every call site pinned to Sora and list the durations and aspect ratios each one asks for.
3. Map durations. Decide per call site whether it becomes a longer Seedance clip or two shorter Veo shots. This is the only step that needs judgement.
4. Re-render one reference shot per call site on both replacements and compare against the Sora output you kept. Prompts that leaned on Sora's look usually need a sentence about lighting or lens, not a rewrite.
5. Set resolution explicitly everywhere, now that it is a parameter rather than a constant.
6. Turn off native audio where you already have a voiceover stage.

If you would rather not run that migration by hand, the same models are one dropdown apart inside Imgo’s video generator, and the same prompt can be sent to both before you commit a pipeline to either.