The Sora API is gone, and OpenAI named no replacement.

OpenAI keeps a table of every model it has retired from its API. Each row has three columns: the shutdown date, the model, and the recommended replacement. The rows for the Videos API, sora-2, sora-2-pro, and their dated snapshots all read 2026-09-24 in the first column. In the third column there is a dash. [1]

That dash is the story. OpenAI told developers on March 24, 2026 that the Sora 2 models and the Videos API would leave the API on September 24, 2026. [1] It kept its word, gave half a year of notice, and pointed nobody anywhere. If you built a creative pipeline on Sora, the migration was your problem from the day the notice arrived.

What OpenAI said, and what it did not

The consumer product went first. OpenAI’s help center says the Sora web and app experiences were discontinued on April 26, 2026, and the API on September 24, 2026. [2] You can export what you made from sora.chatgpt.com/sunset. If OpenAI offers a final export window, it will email you first, and after that window it will permanently delete any data associated with your use of Sora. [2] Purchased credits survive, but only as credits for Codex. [2]

What the page does not contain is a reason, a successor, or a promise about anything besides the media files. That matters more than it sounds. A generated clip is the least valuable thing a production pipeline holds. The prompts that worked, the reference images, the seeds and settings, and the record of which variant won the test are what took months to build, and none of that is guaranteed to come out of an export button.

Sora was not the only one

It would be comfortable to read this as an OpenAI problem. Adobe’s help center suggests otherwise. It keeps a list of the partner models it has retired from its apps, and says models may be removed for low usage, quality concerns, contract expiry, or the end of provider support. [3] Between April and September 2026 that list gained ten entries. [3]

The model most teams will reach for next has a date on it too. On Google Cloud, the generally available veo-3.1-generate-001 lists a release date of November 17, 2025 and a retirement date of November 17, 2026 or later. [4] On the Gemini API, the Veo 3.1 model IDs on the pricing page all end in -preview. [5] Neither is a warning that Veo is going away. Both are a reminder that a video model is a versioned dependency with a published end of life, not a platform.

What the alternatives cost

Pricing is the one part of a migration you can plan precisely, because the vendors publish it per second of output.

Two details in those pages are worth more than the rates. Google says you are only charged if a video is successfully generated. [5] And Runway, which now resells models from several labs next to its own, bills a routed generation at the rate of whichever model its router selects and reports the model used and the realized cost in the response. [6] If you route, log that field. Otherwise you will not know which model made the ad that won.

The cost of a test cycle is simple arithmetic, and it is worth running before you decide how many variants creative testing actually needs.

At the Standard rate, 30 eight-second clips cost $96 a round. On Fast at 1080p the same round is $28.80. The gap between tiers is bigger than the gap between vendors, which is one more reason not to marry either.

Build the pipeline to outlive the model

The teams that lost the least on September 24 were the ones that treated Sora as a replaceable part from the start. The pattern is not complicated, and none of it depends on which vendor you pick next.

• Store every prompt, reference image, parameter, model ID, and output in storage you control, the day it is generated. The Sora deletion clause is the reason. [2]
• Put one internal interface in front of every video provider, so a model swap is a configuration change and not a rewrite.
• Keep a fixed set of test prompts with approved outputs, and score any replacement against it before production traffic moves. That is an eval, and it is the only honest way to say the new model is as good.
• Have a second provider generating real work before the first one sends a notice.
• Record the model that actually made each clip, especially when a router picks it. [6]
• Treat preview model IDs as temporary, and read each vendor’s deprecation page on a calendar, not when a campaign breaks. [4][5]

None of this makes Sora’s shutdown painless for the teams who built on it. It makes the next one boring, and in this market there will be a next one. The column in OpenAI’s table told you the most useful thing about generative video in 2026: the vendor owes you notice, and it does not owe you a path.

Sources

1. OpenAI · DeprecationsThe March 24, 2026 notice, the September 24, 2026 shutdown of the Videos API and the Sora 2 models and snapshots, and the empty replacement column.
2. OpenAI Help Center · What to know about the Sora discontinuationThe April 26 and September 24 dates, the export path, permanent deletion, and credits for Codex.
3. Adobe · Retirement of generative AI models in Adobe productsLast updated September 10, 2026. The reasons models are removed and the dated list of retired partner models.
4. Google Cloud · Veo 3.1The generally available veo-3.1-generate-001, released November 17, 2025, with a retirement date of November 17, 2026 or later.
5. Google AI for Developers · Gemini Developer API pricingVeo 3.1 Standard, Fast, and Lite prices per second, the preview model IDs, and billing only for successful generations.
6. Runway · API Pricing & Costs$0.01 per credit, the per-second credit rates, and how routed generations are billed and reported.

MS

Editor-in-chief

Mikołaj Salecki

Writes about media, tech, and AI business for people who actually run digital. Former agency lead. Skeptic of frameworks that read better than they perform.

LinkedIn ↗Facebook ↗salecki.digital ↗

More articles →