LLMSwaps

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

Questions about swaps that published pages can answer

A question is answered here when an app page or a vendor notice settles it. Anything that needs an API key, a probe or an inference from output quality is out of scope. As of 2026-09-12.

Why this register does not probe for engine namesAn API probe or a look at output could suggest which model an app runs. A suggestion cannot be checked by a reader, would go stale the moment a vendor swapped quietly, and would make this register the source of the claim rather than the app. So the register waits for a published sentence.Where would the answer come fromA published pageIt gets answeredQuoted with the addressand the day it was read.A probeDeclinedA reader cannot repeat it,and a quiet swap wouldmake it wrong withoutwarning.Output that looks familiarDeclinedRecognising a style is notevidence about an engine.Coverage traded for verifiability
Fig. 1 Refusing to probe costs coverage and buys the only thing a notice is for: something a reader can open and verify.

Probing is the tempting shortcut and the one this register refuses. Output can suggest which engine produced it, and a suggestion is not a fact a reader can verify by opening a page. The register would also be wrong the moment a vendor changed something quietly, which is precisely the event being tracked.

The questions that survive that rule are about disclosure rather than about engineering: who names an engine, who commits to a date, what a deprecation notice actually promises, and what a team should have in place before one lands.

Answers link to the notices they rest on, and the notices carry the addresses, so a reader can check the original wording rather than this register's summary of it.

1The questions

  • How much notice — Applications name no figure at all.
  • Your stored work — One vendor published what it would do with stored work when a model shut down.
  • Notice clauses — One pricing page says outright that the models it offers may change, and nobody anywhere publishes how much notice would come with it.
  • Aggregators — Two apps resell models they did not make, which means their menus can change for reasons entirely outside their own control.
  • One or many — One app names a single house model on every plan while another sells two generations at once, and both are stable in different ways.
  • Naming versions — An app that prints the model and version beside the rate makes a retirement visible in advance.
  • Blank dates — An empty calendar is a finding about announcements, not an assurance about schedules.
  • Who names a successor — One dated row here names what follows it, and the successor is a later version of the same model.
  • After the date — One platform names the exact response a retired model returns.
  • Can a date move — One platform states its dates cannot be pushed back.
  • Cut off early — One platform can end access after fifteen days of disuse.
  • Who moves you — One platform states migration will not happen automatically.
  • Published states — Three on one platform, two stages on another, and none anywhere else.
  • Older generations — Three entries list an earlier generation beside the current one.
  • Quiet about the last version — Three entries promote a numbered generation and say nothing about the ones it replaced.
  • Per-model metering — Three entries attach a price to each named model rather than bundling them, so the invoice records which engine produced what.
  • House against licensed — Almost none.
  • Downloadable weights — A copy already downloaded cannot be withdrawn, so a retirement date stops applying and hardware, storage and retention start.
  • Tier and model — One entry names partner models tier by tier, so the subscription decides.
  • Who publishes a date — Six of the sixteen entries show an end date, all of them published by three companies, and none of the three is an application a production buys from.
  • After a withdrawal — Two hosting platforms fail the request once a model is gone.
  • Finished work — Three answers are published: data deleted after an export window, renders kept with their tier recorded, or model weights you hold yourself.

Where nothing has been published, the answer names the absence instead of estimating around it.

Other notices: Apps, Models, Fields, Side by side, Terms, Learn, Data. Where each line comes from is set out on the sourcing page.

2A pipeline across model versions

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

When you edit a sheet (change the hair, swap the coat, add a scar) or generate a state (the drawn blade, the burnt room), the model starts from the existing image and reads your text as what to change.

SceneMixer, Editing a Sheet or Making a State: Write the Change, Not the Whole Picture (read 2026-10-13)

Recommended is the badge on the tier the site suggests for most character sheets; the default for new accounts stays the cheaper tier.

SceneMixer, Qwen-Image 3.0 Pro: the recommended tier — “Is Recommended the same as default?” (read 2026-10-13)