If anyone built Amazon video for you on the Sora API — Sponsored Brands Video variants, a PDP clip, an "AI-powered video" line on an agency invoice — that pipeline stops working on Thursday, September 24. Ten days from today. Inside the creative freeze, two weeks before the October event, and five weeks before the BFCM deal window closes.
The videos already live on your listings don't go anywhere. What goes away is the ability to make the next one in the same model: the Black Friday recut, the new-offer version, the variant for the SKU you add in November. And OpenAI's own deprecation table has nothing in the replacement column. There is no OpenAI video model to move to.
That makes this a different kind of retirement than the ones I've written about all year. It isn't a model migration. It's a vendor change, and a vendor change nine weeks from peak is a decision, not a find-and-replace.
What happened
OpenAI notified developers on March 24, 2026 that the Videos API and all Sora 2 models — sora-2, sora-2-pro, and the dated snapshots sora-2-2025-10-06, sora-2-2025-12-08 and sora-2-pro-2025-10-06 — will be removed on September 24, 2026, following the consumer Sora app shutdown on April 26 (OpenAI Help Center; The Decoder). The recommended-replacement field is empty.
Why most brand owners will read this wrong
Dumb take one: "Sora's dead, so AI video was hype." No. Video generation got dramatically cheaper and more controllable this year — I wrote about Gemini Omni 1.1 Flash's keyframing and per-second pricing three weeks ago. One company leaving a category tells you about that company's priorities, not about whether a 30-second Sponsored Brands Video can be drafted for a few dollars. It still can.
Dumb take two: "Just swap the endpoint." This is the expensive one, and it's the one you'll hear from whoever built the pipeline. Several aggregators advertise an OpenAI-compatible video endpoint where migration is "two lines — base URL and model name." That's true about the code. It's false about the output. A different model interprets the same prompt differently, handles reference images differently, drifts differently across a clip, and renders your label text differently. The job runs, the file arrives, and it's a different creative decision that nobody made on purpose. A loud failure on the 24th would honestly be the better outcome.
I've spent five months telling operators that a silent model swap is the risk that doesn't show up in a dashboard. This is the same risk wearing a deadline.
What actually changes for a brand doing $200K a month
Cost is not the story. Sora 2 API pricing ran roughly $0.10 to $0.70 per second depending on model and resolution. At the top of that range a 30-second SBV was about $21. On the Omni 1.1 Flash pricing I worked through in August, the same 30 seconds at 1080p is about $4.50. Either way it's a rounding error against a $30K monthly ad budget. The cost of video was never the generation. It's the brief, the QA, and the person who checks the label at second 27.
The series problem is the real one. Most brands running AI video on Amazon don't have one clip. They have a family: a hero SBV, two or three hook variants, a PDP cut, maybe a seasonal version. Those clips share a look because they came from one model. A November recut generated on a different model will not match the three clips it runs next to, and at thumbnail size in a muted autoplay slot, a visible style break reads as a different brand. You can't extend a series in a model that no longer exists.
Your October video tests lose their baseline. If you're mid-test on SBV variants, a replacement clip from a new model changes two variables at once — the message and the model. You'll get a number and you won't know what it measures.
Two video fuses inside six days. On September 6 I flagged that gemini-omni-flash-preview is deprecated September 30. If your vendor moved off Sora onto the Omni preview earlier this year — a reasonable choice in May — they are about to hit a second retirement six days after the first. Ask which string they're on, not which company.
The disclosure question resets. If any of your AI video contains a generated person — including hands — New York's synthetic performer disclosure law has applied since June 9, and Amazon's July rule requires the metadata flag. Regenerating on a new model is a new asset. Whoever makes it needs to run the disclosure and metadata check again rather than assuming it carried over from the old file.
The notice probably didn't go to you. Deprecation emails go to the owner of the OpenAI organization that made the API calls. On a lot of $200K/mo brands, that org belongs to an agency, a freelancer, or a developer who set it up in 2025. You may be the last person in the chain to find out.
What I'd do this week
1. Send one email to every creative vendor. "Does any video asset or pipeline on our account use the OpenAI Sora API or a Gemini Omni preview model? If yes, which assets, and what is your plan for September 24 and September 30?" A shop that's on top of it replies with a list and two dates. A reply that's a paragraph about their AI-powered creative process means nobody has looked, and that reply is your answer.
2. Archive everything by the 23rd. Every final render at full resolution, every prompt, every reference image, every keyframe, in a dated folder you control. Don't assume anything a vendor didn't download and store will be retrievable after shutdown. Prompts are the part people forget, and they're the only record of why a clip looks the way it does.
3. For Q4, recut instead of regenerating. You already own the masters. A new offer line, a different end card, a shorter hook, a trimmed version for a new placement — that's an editor and an afternoon, not a model. It keeps the series consistent, it's measurable against your existing variants, and it doesn't put a new vendor into your peak creative.
4. If you truly need new generation before January, pick one model and re-run the golden set. Six clips from six real SKUs, freeze-frames at 0, 10, 20 and 30 seconds, and read every label and every number. Pin the exact model string, not a family name or an aggregator alias. Draft at low resolution, judge at 720p, deliver at 1080p. Ship on Sponsored Brands Video first, where a CTR read comes in two to three weeks, not on the detail page.
5. Add two columns to the automation inventory. You already have model string, pinned or floating, and retirement date. Add "replacement: named / none" and "which assets on Amazon came from this." A retirement with a named successor is a migration. A retirement with a blank successor is a sourcing decision, and it needs a different owner and a lot more lead time.
What I'd ignore
The Sora economics gossip. There are figures circulating about Sora's daily operating cost and lifetime revenue. I can't source them to anything OpenAI published, and even if they're right they don't change a single frame of your SBV.
"OpenAI has abandoned video" strategy takes. Maybe. It's a capital allocation story about a company you don't run. Your question is narrower: which of your assets depend on a string that stops resolving on the 24th.
Alternative-model leaderboards. No published benchmark measures whether your label survives second 27 of a muted autoplay clip at thumbnail size. You still build that eval yourself, from your own SKUs.
"Drop-in replacement" pitches. Anything sold as a seamless swap is selling you the exact failure mode above — the code keeps running and the creative quietly changes.
The urge to rebuild your video strategy before Christmas. Nine weeks from peak is the wrong moment to add a new vendor to your creative pipeline. Recut what you own, archive what you made, and run the migration in January when a bad clip costs you a Tuesday and not a Cyber Monday.
Every retirement I've written about this year had a successor model and a migration guide. This one has a shutdown date and a blank field. The live clips are fine. What you're losing is the ability to make a matching one, and the fix is a folder, an email, and choosing not to generate anything new until January.