Lifecycle state: what a model is allowed to be

A state is a published label a model sits in while it is on its way out. Amazon Bedrock names three and exposes them through its own interface, Microsoft Foundry separates preview from generally available, and Google either lists a model with a shutdown date or does not list it. No application names a state at all. From 4 source pages, read 2026-09-12 to 2026-09-22; no maker pays to be listed.

Inclusion rule. A state counts when the entry publishes the label itself, not when a reader could infer one from a product going quiet. Marketing words like new and classic are not lifecycle states. Order. Entries publishing a state first, then the applications as one row.

1A state is what turns a removal into a process

Without a published state, a model has exactly two conditions from outside: it answers, or it does not. Everything a production would want to plan around sits in between those two, and a vendor that names the in-between has given a reader somewhere to stand.

That is why the platforms publish states and the applications do not. A platform is talking to somebody whose software has to branch on the answer; an application is talking to somebody picking a tier, and a tier that admits to having a middle condition looks broken rather than honest.

2One entry lets you read the state instead of the page

Amazon Bedrock puts its three states on the model itself, so the answer arrives through the same call that runs the work. That is a materially different thing from a table a human has to remember to re-read, because it can be checked on a schedule nobody has to keep.

It also means the state is checkable at the moment of use rather than at the moment somebody thought to look. A model that moved to legacy last Tuesday is discoverable on Tuesday, which is the only version of this information that survives contact with a busy production.

3What the blank rows here are not

Twelve entries in this register name no state. Read plainly that means nothing about their intentions and everything about what a reader can verify: there is no published condition for a model to be in, so there is nothing to watch changing.

The practical substitute is the menu itself. A model that appears on a priced page one month and not the next has moved through every state at once, unannounced, and the only evidence is two dated readings of the same page. That is what the calendar and the data file are for.

4This column, entry by entry

The lifecycle states column read one entry at a time, wherever that entry has published something here or falls silent in a way the others do not. Ordinary blanks stay on the table above.

• Amazon Bedrock — three states, published on the model and readable through the interface
• Microsoft Foundry — preview and generally available are states, and which one applies decides the notice owed
• Gemini API — being listed on the deprecations table is itself the state, and the date is separate
• Sora — discontinued, announced once and applied to two surfaces five months apart
• LTX Studio — no state vocabulary at all, and a clause where the states would have been
• Kling AI — current and previous generation function as states without being published as any
• Wan-AI — a repository is published or it is not, and a copy already taken has no state at all

• The three states a model can be in
  Active, Legacy or End-of-Life, readable from the modelLifecycle fieldAmazon Bedrock, model lifecycle / recorded 2026-09-22
• What the endpoint does after the retirement date
  All inference requests return 410 GoneMicrosoft Foundry, lifecycle and support policy / recorded 2026-09-22
• What a date in that table commits to
  The shutdown dates listed indicate the earliest possible dates on which a model might be retiredGoogle, Gemini deprecations / recorded 2026-09-22
• Sora web and app
  26 April 2026discontinuedOpenAI Help Center / recorded 2026-09-12

5When the model underneath changes

Model end dates touch “Lifecycle states”; SceneMixer's guides describe how its pipeline treats the models it runs:

SceneMixer turns a novel into a watchable short-drama series in one pipeline: upload the novel (Chinese or English) → AI extracts characters, scenes, and an episode-by-episode structure → generates consistent character and scene art → writes storyboard scripts → renders video segment by segment (16:9 by default, 9:16 in one setting) → composites full episodes.

SceneMixer, Turn a Novel into a Video Series with AI (read 2026-10-13)

Seedance 2.0 (ByteDance's video model family) is one of the strongest engines for vertical short drama in 2026.

SceneMixer, Seedance 2.0 in a Real Short-Drama Pipeline (read 2026-10-13)

6Sources

Values come from the page the entry is read on: vendor announcements for endings, app pages for menus, both linked from the calendar. What counts as a value is on sourcing. Read 2026-09-12. Other fields: Ending early, Named replacement. All of them: the field list.