LLMSwaps

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

Every notice in one file, with its date

The notices are published as one file. Each row records what an app or a vendor stated, where it was stated, and when this site read it. As of 2026-09-12.

Why the file is a series rather than a snapshotA single reading says which model an app names today. Two readings months apart are what a swap looks like from outside the company. The reading date on every row is what turns the file from a directory into evidence of change.First readingThe app names a modelon a public page.Recorded with thedate.Months passSecond readingThe same page namessomething else, ornames nothing.CompareA swapTwo dated rows withdifferent values. Theevent, not a rumourabout it.
Fig. 1 Nothing here is inferred from output, so the series is thin and the thin parts can be trusted.

This register is about change, so the file is mostly useful as a series. Two readings of the same app months apart are what a swap looks like from outside, and a single snapshot cannot show it.

Nothing inferred is exported. If an app names no engine, the file says the name is not published rather than guessing from output, which means the data is thinner and can be trusted where it is not thin.

Dates are split across their own columns for the same reason. An announcement date, a deprecation date and a shutdown date are three different commitments, and collapsing them into one field is how planning goes wrong.

The licence below places the data in the public domain, so no permission is needed. Keeping the source column anyway matters more here than most places, because the claims in it are about other companies.

1What each column holds

Reading one row of the notices export. Recorded 2026-09-12.
ColumnWhat it holds
fact_idThe notice's reference. It stays put when a date is revised.
valueThe commitment as the vendor published it.
applies_toWhether the row is about an app, a model or a date.
source_urlThe changelog or deprecation page it came from.
source_nameThe vendor's heading for that notice.
value_sinceWhen the commitment was first visible.
verifiedThe last reading of that notice.

Inclusion rule. Every notice is exported, including rows recording that an app names no model at all. Order. File order follows recording order, grouping rows by app and then by model.

The file is at /datasets/swaps.csv, licensed CC0 1.0. No attribution is required, though naming the reading date keeps a quoted figure checkable.

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

2Production that outlasts a model

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

Each scene gets an automatic AI 3D model, and every shot in the scene is generated against it.

SceneMixer, Why a Whole Series Stays Consistent on SceneMixer — “Does the layout of a room really stay the same across shots?” (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)

The quote column. Each reading in the file also carries the words on its source page that it was read from, so any row can be checked against that page. Rows recorded as not stated on the pages read leave it empty.