On September 24, 2026, OpenAI removed the Sora 2 video models and the Videos API from its API. Any request to sora-2 or sora-2-pro now has nowhere to go. This site is independent and is not affiliated with OpenAI; Sora is an OpenAI trademark, and the product is closed. The site's old early access page explains why no waitlist can open it again.
The short version
- Removed: the Videos API and every Sora 2 model name, including the dated snapshots.
- OpenAI names no replacement. The "recommended replacement" column is empty for every Sora row.
- There is no drop-in swap. Another vendor means a new request shape, new polling and new output handling.
- Microsoft's hosted copy of Sora 2 runs out on its own schedule, with the last version set to retire on October 15, 2026.
What was there before
The Videos API was how code talked to Sora. You posted a prompt to the v1/videos endpoint with a model name, waited for the job, then downloaded the file. The model page says the Videos endpoint was the only one the Sora 2 models supported: no Responses, no Chat Completions, no Image endpoint.
The announcement came earlier. OpenAI's deprecations page dates the notice to March 24, 2026 and the removal to September 24, 2026, so developers had six months of warning.
What was switched off
| Item | Status as of October 15, 2026 |
|---|---|
Videos API (v1/videos) | Removed |
sora-2 (alias) | Removed |
sora-2-pro | Removed |
sora-2-2025-10-06 | Removed |
sora-2-2025-12-08 | Removed |
sora-2-pro-2025-10-06 | Removed |
| Recommended replacement | None listed |
Source: OpenAI API deprecations page, entry "2026-03-24: Sora 2 video generation models and Videos API".
Why this matters for video work: the Sora 2 models supported only the Videos endpoint, so there is no other OpenAI call that still reaches them. If a pipeline mixed text, images and video from one vendor, the video step is the one that has to move.
What to change in code
Start with a search for the model names and the /v1/videos path. Those are the strings that will fail. Then look at what sat around the call:
- the job polling loop, because another service will have different job states and different timeouts;
- the download step, since output URLs, formats and expiry are the provider's choice;
- prompt text written for Sora's habits, which will not carry over unchanged;
- any cost estimate built on Sora's per-second pricing.
A hosted alternative is one route. Google documents Veo in the Gemini API on its video generation page (see the Gemini API docs under sources). That is a different request shape, so budget for rewriting the call rather than renaming a model. Open-weight video models run through ComfyUI are the other route: no vendor can remove them, but they need your own GPU and setup time.
Microsoft's copy of Sora 2
Teams that used Sora 2 through Microsoft Foundry were on a separate clock. Microsoft's retirement schedule lists sora-2 version 2025-10-06 as retiring on July 15, 2026, pointing to version 2025-12-08, and that version as retiring on October 15, 2026, with no replacement named. The page was last updated in September 2026, so check it before relying on the date.
What is still unclear
OpenAI's pages do not say whether a successor video API is planned. Absent that, treat any claim about a "Sora replacement" as a product comparison, not a migration path.
Who should move now: anyone whose app still calls the old endpoint, since it fails today. Who can wait: people who only used the app for personal clips, as the API change does not touch them.