The OpenAI Videos API and every Sora 2 model are removed on 24 September

The OpenAI Videos API and every Sora 2 model are removed on 24 September

22 September 2026 — OpenAI's deprecations page records that on 24 March 2026 it notified developers using the Videos API and the Sora 2 video generation aliases and snapshots that all of them would be removed from the API on 24 September 2026. That is tomorrow, counted from the day this was written.

Six months' notice is generous by the standards of this market. It is also long enough that whoever read the email in March may no longer own the code.

What else stops answering this quarter

The Videos API is not the only entry on that page with a date inside the next twelve weeks. Read in order, the queue is:

• 24 September 2026 — the Videos API and every Sora 2 alias and snapshot.
• 28 September 2026 — four legacy model snapshots announced a year earlier, replaced by a current text model.
• 1 October 2026 — a specialist security-focused model, announced on 11 September, replaced by its predecessor generation's equivalent.
• 23 October 2026 — twelve legacy snapshots and their fine-tuned derivatives, announced in April.
• 30 November 2026 — the reusable prompts API, the evaluations dashboard and API, and the agent builder, announced in June. This one is different in kind: it is tooling, not a model, and there is no drop-in replacement of OpenAI's own.
• 1 December 2026 and 11 December 2026 — an image model generation and six more snapshots.

Why the fine-tuned derivatives are the trap

A fine-tuned model is not a separate thing that survives its base. When a base snapshot is retired, the models you trained on it stop answering with it, and the training run — and whatever internal approval went with it — has to be repeated against a current base. Teams budget for the migration of a prompt and forget the retraining, the re-evaluation and, in regulated settings, the re-documentation.

The 30 November entry deserves separate attention for the opposite reason. Prompts, evaluations and an agent builder are the things a team builds its own process around. Replacing a model id is a code change; replacing an evaluation harness is a change to how you decide whether anything works.

What to do about it

Export the model ids your account actually called in the past ninety days rather than the ones you think it calls — usage exports exist for this — and put the list beside the deprecations page. Anything that matches a dated row is work with a deadline on it.

Then write the dates into whatever your team actually reads: a calendar, a ticket, a renewal tracker. The failure mode with model deprecations is never that the notice was too short. It is that six months of notice produced no artefact anybody would trip over in month five.

Not yet independently verified. This rests entirely on OpenAI’s own deprecations page, including the March notification date and every shutdown date listed. No independent report of it was found, so every date here comes from the vendor's own page and is only as reliable as that page. We will update this when it can be confirmed, and remove this note.

Sources

• OpenAI API deprecations page — 22 September 2026