The Sora API Is Gone: OpenAI's Quiet Exit From Video

OpenAI's Videos API and the sora-2 model family stopped answering on September 24, 2026. Since the next day, calls return 410 Gone. That status code is worth pausing on. A 500 means the server stumbled and you can try again. A 410 means the thing is gone for good, and no amount of retrying will bring it back. If your job queue still has a retry loop pointed at that endpoint, it is burning cycles on an answer that will never change.

The quieter detail sits in OpenAI's own deprecations page. The column for a recommended replacement is a dash. For every Sora model, the suggested successor is nothing at all. The company has not announced a Sora 3 or a different video model. As of early October, OpenAI simply does not sell video generation.

Two shutdowns, nine months apart

The end came in two stages, and the dates matter if you are trying to work out what happened to your old projects. Sora first appeared as a research preview on February 15, 2024, shown to red teamers and a small group of artists and filmmakers. On December 9, 2024, it opened to ChatGPT Plus and Pro subscribers at sora.com. Sora 2 followed on September 30, 2025, alongside an invite-only iOS app in the United States and Canada. That version added synchronized audio, a feed for discovering and remixing clips, and a feature for dropping yourself and your friends into scenes. Clips ran up to 15 seconds, or 25 on the web for Pro users. Disney signed a three-year licensing deal in December 2025 to bring more than 200 characters to Sora, and agreed to invest a billion dollars in OpenAI.

Then the retreat. On March 24, 2026, OpenAI told users the app and the API would shut down on separate dates. The consumer app and web experience closed on April 26. The API and the Sora 2 snapshots were removed on September 24. Users were pointed to an export page and told that once the final export window passed, OpenAI would permanently delete the data tied to their Sora usage. The company said the Sora research team would shift toward world simulation work, aimed at robotics.

What a 410 actually costs you

Most teams do not feel a shutdown on the day it happens. They feel it weeks later, when a render queue turns red and nobody remembers which brief was supposed to run on which model. The awkward part of the Sora exit is that there was no migration path handed to developers. When a vendor deprecates a model and names a replacement, you can usually script the swap and move on. Here the replacement column was empty, so every team has had to re-brief its backlog by hand.

The practical move is to sort the old Sora jobs by what they actually needed rather than by how they looked. If a shot required someone talking on camera, that is a job for a model with native audio and tight lip sync. If it required one long take where a face has to hold together, that pushes you toward models that generate longer single shots. If it required volume with sound at a low price, the choice is different again. Wrapping the old endpoint in a silent fallback is the tempting shortcut, and it is also the one that lets a client's brief quietly run on a model they never approved.

Why OpenAI might walk away from consumer video

Standing back, the decision fits a pattern that has been building for a while. Consumer video generation is expensive to serve and hard to defend. The per-second cost of inference is high, the outputs are heavy to store and deliver, and the legal surface is wide: likeness, copyright, and the licensing deals that keep studios comfortable all add friction. Meanwhile the same research talent can be pointed at simulation problems that feed robotics, which is a market OpenAI has said it wants to serve.

There is also a competitive story. While OpenAI stepped back, everyone else stepped forward. Google pushed its Omni Flash line to general availability in late August, with clips of three to ten seconds, native audio, and output up to 4K. ByteDance's Seedance family kept extending single-shot length. Chinese labs like Shengshu have been dropping prices to fractions of a cent per second. In that field, a premium, consumer-facing video product is a harder sell than it was two years ago.

The part nobody should skip

The lesson is not really about Sora. It is about what happens when a hosted model becomes a dependency. A model you rent can be retired on a date you do not control, and the announcement may not include a replacement. That is true for video, and it is just as true for text, image, and audio models that teams have wired into production without a fallback plan.

Three habits make the difference between a small inconvenience and a broken pipeline. Keep the briefs, not just the outputs, so a shot can be re-run on a different model without starting from scratch. Avoid silent fallbacks that hide which model produced which asset, because that erases your ability to explain a result later. And treat the deprecation page as a first-class engineering input, not a help-center page someone reads after the alerts start firing.

What the exit says about the video market

Look at who stayed and the picture sharpens. The companies still investing in video generation tend to be the ones for whom video is not a side feature. Google has an advertising and entertainment business that benefits directly from better generated video. ByteDance owns short-video platforms where generated clips can be used immediately. Those businesses can absorb the cost of serving video inference because the output feeds a product they already monetize. For a lab whose main revenue comes from chat subscriptions and enterprise API calls, consumer video is a much thinner proposition.

That logic explains the shapes of the two exits. The Sora app closed first, in April, because a consumer app carries the heaviest obligations: moderation, storage, delivery bandwidth, and a support burden that grows with every viral clip. The API followed in September, once the team had decided not to offer a successor. The gap between the two dates tells you the retreat was planned as a wind-down rather than a surprise reversal.

For developers, the practical read is that video generation is consolidating around a smaller set of providers, most of them attached to a platform. That is not necessarily bad. It does mean that anyone building a video product should think about the second-order risk of depending on a single vendor, and should keep at least one open-weights path in reserve, even if it currently produces worse output.

Sora was, for a while, the most visible name in AI video. Its exit is a reminder that visibility and durability are different things. The models that survive are the ones a business can actually plan around, and increasingly that means open weights you can run yourself or a vendor with a clear road map. Everything else is a rental, and rentals end.