How much warning anyone actually gives
It depends entirely on who you buy from. No application tracked here publishes a figure. Amazon Bedrock commits to six months or 45 days depending on the model, Microsoft Foundry to at least 60 days for a generally available model and 30 for a preview one, and the Gemini API promises advance notice without giving a number. From 4 source pages, read 2026-09-12 to 2026-09-22; no maker pays to be listed.
| Entry | What is published | Detail |
|---|---|---|
| Amazon Bedrock | Six months or 45 days | Stated per model, before anything is scheduled |
| Microsoft Foundry | At least 60 days, 30 for a preview | Stated as policy, not extendable |
| Gemini API | Advance notice, uncounted | Promised in words, never in days |
| Sora | Nothing promised | About five months elapsed between two closures |
| Every application tracked | Nothing published | No figure in either direction |
Inclusion rule. A figure counts when the company that makes the model, or the platform that serves it, states it on its own page. An interval observed after a retirement is not a commitment and is kept in the detail column. Order. Entries stating a figure first, longest first, then the rest.
1The answer splits by who is selling, not by how big they are
Read across the register and the pattern is clean: a notice period appears wherever the buyer is a developer who wrote a model identifier into code, and vanishes wherever the buyer is a creative team choosing a tier. The cloud platforms publish numbers because their customers will file support tickets quoting them. The applications publish none, and the productions that use them have nothing to quote.
That also means the useful number is the one belonging to the party you actually call. A production generating through an application inherits none of the platform commitments below it, even when the model it is using is the same model, hosted by the same company, with a published date attached. The commitment stops at whoever holds the contract.
2The two surfaces closed in the order that matters
Consumers lost access first and developers later. That ordering is the friendlier one: the people with pipelines, contracts and scheduled work got the longer runway. The reverse ordering would have left production teams stranded while casual users carried on.
Nothing published says the ordering was deliberate. It is recorded here because it is the shape of the only consumer-product closure this register has watched, and a second one will either match it or make it look like an accident.
3What a register can say about the empty rows
Every other model here has no announced end date. That is a statement about announcements and not about intentions: a vendor with no plans to retire anything publishes exactly the same blank as a vendor two weeks from an announcement.
Reading a blank as safety is the mistake this whole register exists to prevent. There is also a trap inside a filled cell. One platform states that once a model enters its legacy period, an existing customer may lose access after fifteen days of inactivity, which means a series that pauses between episodes can be cut off without any published date having passed.
- Sora web and app26 April 2026discontinued
- Sora API24 September 2026discontinued
- Notice for a generally available modelAt least 60 days before retirementsent to subscriptions with active deployments
- Notice for a preview modelAt least 30 days before retirementwhich is the stage sora-2 sits in
- Legacy periodSix months or 45 days, stated on each model cardwhich is the notice given before end-of-life
- How access can end before the dateOnce the Legacy period begins, existing customers may lose access after 15 days of inactivity
- How many days of advance noticeNo number of days published (as of 2026-09-12)promised but never counted
4Production that outlasts a model
Model end dates touch “How much notice”; SceneMixer's guides describe how its pipeline treats the models it runs:
Dubbing replaces the audio of a finished clip; the picture was generated for a different line, so mouth shapes, pauses and shot length all belong to the original language. Native-language generation gives the video model the line in the target language before it renders the shot, so the performance, the pacing and the cut points are built for that language.
SceneMixer, Native-Language Dialogue vs. Dubbing in AI Video (2026): Why the Cast Should Speak the Line (read 2026-10-13)
Wan 3.0 is Alibaba Cloud's video generation model on the Model Studio platform, called through an asynchronous submit-then-poll API, in two tiers: the standard wan3.0-video and the faster wan3.0-video-prime.
SceneMixer, Wan 3.0 for AI Short Drama (2026): Parameters, Reference Limits, Moderation Rules and the 30-Second Take (read 2026-10-13)
5Sources
Dates come from the company that makes the model or the platform that serves it; menus come from the app, and both are linked on the calendar. Read 2026-09-22. What counts as a date is on sourcing. Nearby: Your stored work, Notice clauses, Aggregators.