The shutdown date has passed, and the replacement column is empty
OpenAI switched off the Sora API on 24 September. The company’s own help centre notice states it plainly: “The Sora web and app experiences were discontinued on April 26, 2026. The Sora API will be discontinued on September 24, 2026.”
The deprecations page lists what went with it: 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, together with the Videos API itself. The deprecation was announced on 24 March, giving six months’ notice.
The column that usually names a migration target is blank. OpenAI has deprecated a lot this year — gpt-5.4-cyber announced on 11 September for shutdown on 1 October, the transcription models including whisper-1 announced in August for February, legacy audio and realtime families announced in July for January, GPT-5 and o3 snapshots announced in June for December — but those point somewhere. This one does not, because there is no OpenAI video model left to point at.
Why an empty replacement column is more than an inconvenience
A normal deprecation is a string change. You swap gpt-5-2025-xx-xx for its successor, re-run your evaluations, and ship. The work is real but bounded, and the output format stays roughly the same.
A deprecation with no successor is a different job. It forces a vendor evaluation from scratch: different request and response shapes, different parameters, different safety behaviour, different pricing, different latency, and a new contract. Anyone who wired Sora calls into a product, a plugin or a live service has been doing that work rather than a rename, and other video models — Kuaishou’s Kling and Luma’s Ray 2 among them — sit behind their own APIs with their own semantics.
The six-month notice period is the mitigating fact here, and it is worth stating plainly: OpenAI did not pull this without warning. What it did not do was offer anywhere to go.
What OpenAI has said, and what it has not
The help centre notice explains the mechanics and not the reasoning. It tells users they can export their Sora content at sora.chatgpt.com/sunset, that a final export window may be offered with email notice, and that after that window “we will permanently delete any data associated with your use of Sora.” Purchased ChatGPT and Sora credits, it adds, can still be used for Codex.
Nowhere in the notice does OpenAI say why. That absence is itself informative in a year when the company has paused training and tool-using inference on its most capable models and directed research attention toward coding and agents. Sora launched as a ChatGPT feature rather than as an API-first product, and it competed for the same compute as everything else. Neither of those is an explanation OpenAI has offered, and neither should be reported as one.
What to watch
The immediate question is whether the credit transfer to Codex holds for everyone who bought them, and whether the final export window materialises before deletion begins. The help centre hedges it — “if we are able to offer one” — which is unusual phrasing for a company sunsetting a paid product, and worth following.
The larger question is what OpenAI’s product map for video looks like now. The company has a text-to-video capability inside its research history and no shipping video endpoint. Whether that is a permanent exit from generative video or a pause before something API-first returns is not something the deprecation notice answers, and OpenAI has not said.
For developers, the practical advice is unglamorous: check whether any code path still calls the Videos API, because it now fails rather than degrades, and export anything still sitting in a Sora account while the export route is open.
