LLMSwaps

Announced end dates for video models, and which apps still carry them

Inference: the request a retirement stops

Inference is running a model to produce output. It is the unit a retirement applies to: one platform here states that after a date all inference requests return a specific error. It is also the unit most pricing in this register is measured in, per second or per finished clip. As of 2026-09-22.

What a retirement does to the different things a request involves. Recorded 2026-09-22.
Part of a requestWhat a retirement does to it
The weights that answerUnreachable, unless you hold a copy
The output already returnedNothing; it is wherever you put it
The prompt and settingsKept, and no longer reproducible
The billStops, which is the smallest consequence

Inclusion rule. Distinctions drawn from the lifecycle pages and priced pages used as sources in this register. Order. From what is lost to what is not.

1A prompt without its model reproduces nothing

Teams instinctively archive prompts and settings so anything can be remade. In this field the thing that will not exist later is the model, so the archive preserves the instructions for a machine that has gone.

Which inverts the usual advice: keep the delivered files at the quality delivered, and treat the prompt as a note rather than as a backup.

2The billing unit and the lifecycle unit are the same thing

Because pricing is per second or per clip, a model leaving a price list removes a purchasable unit rather than a plan. On one entry here the invoice therefore records which engine produced what, as a side effect.

Where models are bundled into a subscription, the same connection is impossible to reconstruct. That difference decides whether a production can answer the only question a retirement raises.

A prompt without its model reproduces nothingTeams archive prompts and settings so anything can be remade. In this field the thing that will not exist later is the model, so the archive preserves instructions for a machine that has gone.What survives a retirementThe delivered fileSurvives every retirement. The one asset that does not depend ona roadmap.The prompt and settingsSurvive, and become unusable the moment the model does notanswer.The modelThe part that goes. Nothing you hold substitutes for it unlessyou hold the weights.Archive the outputs at the quality delivered, outside the vendor's account
Fig. 1 Which inverts the usual advice: keep the delivered files, and treat the prompt as a note rather than a backup.

3A named failure is better than an unexplained one

One platform states the exact response after its date, so a caller can branch on it and alert on it. Everywhere else the request simply fails and the cause has to be investigated.

That is the difference between a retirement that software handles and one a person discovers. The status term covers what that response means.

A definition rather than a dated entry: no line above is a vendor statement. Where this word turns up in a real notice, the wording and the date are on the reproducibility term. Nearby terms: 410 Gone, Region availability, Inactivity clause.

4A pipeline across model versions

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

Voices and ambience are in the clip, and in the composited episode, with the episode’s score unless you turned the Music switch off.

SceneMixer, Dialogue and Sound in One Generation — “Does the downloaded file include the sound?” (read 2026-10-13)

Image and video models take a description as a list of things to put in the frame.

SceneMixer, Describe the Cause, Not the Result: The One Rule Behind Good Sheets and Good Shots (read 2026-10-13)