LLMSwaps

Which model each app runs today, and when it last changed

Who says what replaces a retired model

Almost nobody. One dated row on one platform names its replacement, a later version of the same model. That successor is itself scheduled to retire and names nothing after it, so the documented chain runs one link. Every other ending in this register arrives with no destination at all. As of 2026-09-12.

What has been announced, entry by entryFilled where the vendor has announced something in that column, hollow where it has not.What has been announced, entry by entryWhat is publishedMicrosoft FoundryMicrosoft Foundry — What is published: A successor named for one dated rowAmazon BedrockAmazon Bedrock — What is published: No successor, and migration is not automaticGemini APIGemini API — What is published: No replacement column on the tableSoraSora — What is published: No replacement is named in the noticeEvery application trackedEvery application tracked — What is published: Nothing published
Fig. 1 Filled where the vendor has announced something in that column, hollow where it has not.
What each entry has published about the model that takes over. Recorded 2026-09-12.
EntryWhat is publishedDetail
Microsoft FoundryA successor named for one dated rowA later version of the same model
Amazon BedrockNo successor, and migration is not automaticStated plainly rather than implied
Gemini APINo replacement column on the tableNewer models documented elsewhere, as products
SoraNo replacement is named in the noticeSo the runway led nowhere in particular
Every application trackedNothing publishedAn older generation on sale is not a successor

Inclusion rule. A successor counts when the notice announcing an ending also names what follows it. A newer version existing elsewhere is not a published statement. Order. The entry naming one first, then the rest.

1A date without a destination is half a plan

The work a runway is for is testing whatever comes next, and that cannot start until somebody decides what next is. So a notice with a date and no successor hands a reader the cheaper half of the problem.

Stopping on time is arithmetic. Choosing a target and re-establishing a look on it is the expensive part, and nothing in most of these announcements reduces it.

A documented chain that runs one linkThe earlier of two dated rows names its replacement, a later version of the same model. That replacement carries its own date and names nothing after it, so the published succession runs one step and stops.A dated rowNames the laterversion of the samemodel as itsreplacement.ThenThat versionCarries its ownretirement date, withno successor besideit.ThenUnwrittenNothing publishednames what follows.The chain ends here.Every other ending in this register arrives with no destinationEven where a successor is named, nothing says the platform performs the move for you.
Fig. 2 Reading only the first row suggests succession is managed; reading both shows one step documented at a time.

2The obvious guess stays a guess

Most readers will assume a maker's newer model is the route, and most will be right. Writing that into the register would make every such cell unverifiable, which is the trade refused in every column here.

The cost is not abstract. A team that migrated to an assumed successor and found the pricing, the interface or the region coverage different has spent its runway on the wrong test.

3Naming is not moving

Even where a successor is named, nothing says the platform performs the change. One platform states outright that migration will not happen automatically, which is more useful than leaving it implied.

A complete disclosure therefore has three parts: a date, a destination, and whose job the move is. No entry in this register publishes all three for the same model.

4Sources

Dates come from the vendor that makes the model; menus come from the app, and both are linked on the calendar. Read 2026-09-12. What counts as a date is on sourcing. Nearby: After the date, Can a date move, Cut off early.