AIStall
EditionBy The AI Stall editorsOctober 16, 2026

The Model Shutdown Calendar

The Model Shutdown Calendar

Models get switched off, and the dates are published months ahead on pages almost nobody reads. OpenAI, Anthropic and Google each keep a deprecation table, each with its own notice policy: six months in one case, at least 60 days in another, and earliest-possible dates in the third. The nearest retirement on the three tables is November 30, 2026, which is seven weeks away. This note collects the dated entries and explains how much weight each date carries. All three pages were read on October 12, 2026.

The nearest dates, in order

ChatGPT Business's public homepage, captured automatically on October 10, 2026. Not a logged-in account; the page may have changed since.
ChatGPT Business's public homepage, captured automatically on October 10, 2026. Not a logged-in account; the page may have changed since.

Anthropic's table also carries tentative retirement dates for active models, which read as floors rather than plans: claude-haiku-4-5-20251001 not sooner than October 15, 2026, and claude-opus-4-5-20251101 not sooner than November 24, 2026.

Those two are the entries to watch. An active model with a floor date a few weeks out is a model that could get a deprecation notice at any time.

Claude Team's public homepage, captured automatically on October 10, 2026. Not a logged-in account; the page may have changed since.
Claude Team's public homepage, captured automatically on October 10, 2026. Not a logged-in account; the page may have changed since.

Anthropic publishes a floor date for every active model

This is the most useful column on any of the three pages, and only Anthropic has it. Active models carry a tentative retirement date phrased as not sooner than a given day, so you can see how much runway a model has before any notice arrives.

The pattern is roughly a year of guaranteed life from release. Haiku 5.5 shipped on October 7, 2026 and its floor is October 7, 2027, which is a usefully explicit commitment for anyone building on it.

The page also notes a limit on its own dates. They apply to the Claude API, Claude Platform on AWS and Microsoft Foundry, while Amazon Bedrock and Google Cloud set their own schedules, so the same model can have two different end dates.

The notice periods are not comparable

OpenAI attaches a notice figure to each announcement. The October 1, 2026 entry retiring gpt-5.1 and two others states six months notice, while the text-to-speech entry from the same day states at least three months.

Anthropic publishes one policy for everything: it notifies customers with active deployments and provides at least 60 days notice before retirement for publicly released models.

Sixty days is the shortest of the three, and the Sonnet 4.5 timeline matches it almost exactly. Deprecation was announced September 30, 2026 and retirement is set for November 30, 2026.

Anthropic also publishes something unusual in its downsides section: it acknowledges that retirement removes models from researchers and carries model welfare risk, and says it has committed to long-term preservation of model weights.

Google publishes earliest dates, not dates

Google's deprecation page is explicit about what its table means. The shutdown dates indicate the earliest possible dates on which a model might be retired, and the exact date is communicated to users with advance notice.

That is a weaker commitment than a fixed date, and it cuts both ways. The model may last longer than the table suggests, but you cannot plan a migration around a date Google has not promised.

Several Gemini models have already gone. The models page marks Gemini 2.0 Flash, Gemini 2.0 Flash-Lite, Gemini 3 Pro Preview, Gemini 3.1 Flash-Lite Preview, Imagen 4 and Computer Use as shut down.

There is also a quieter restriction worth noting. The same page says access to the 2.5 models is being limited to users who have actively used them in the past, which is retirement for new projects without a retirement date.

Parameters retire too, and that one returns an error

Anthropic's page lists an API parameter deprecation most teams will meet before any model shutdown. The temperature, top_p and top_k parameters return a 400 error when set to a non-default value on Claude 4.7 and later models.

The Python SDK from version 1.0 removes those parameters entirely, so passing them raises a TypeError rather than failing at the API. The recommended replacement is prompting rather than sampling settings.

That is a breaking change in ordinary code, not in an exotic feature. Any wrapper written a year ago probably sets temperature somewhere.

How to find out whether this is your problem

Both vendors publish a way to check instead of guessing. Anthropic's page gives the steps: open the Usage page in the Claude Console, click Export, and review the CSV to see usage broken down by API key and model.

That five minute check is the whole job for most small teams. Either a retiring model id appears in the export or it does not.

Validation afterwards is cheap if you use the right mechanism. Running an evaluation set at the batch rate costs half, as covered in our note on batch discounts.

Price changes hide inside migrations

A forced move is also a repricing. The replacement for gpt-5.1 is gpt-6-sol, listed at $2.00 per million input tokens against $1.25 for gpt-5.1 on the OpenAI pricing page, checked October 12, 2026.

That is a 60 percent increase on input by our arithmetic, before any caching. The cached rates move too, which is why the cached-input note is worth rereading when a migration lands.

Sometimes the move is cheaper. The replacement for gpt-5.4-nano is gpt-6-luna at $0.10 input and $0.50 output, in the same bracket as Haiku 5.5.

Seat buyers mostly escape this

If your team uses ChatGPT Business or Claude Team rather than an API key, model changes arrive as a quiet swap in a model picker. That is one of the real arguments for buying seats.

The trade is control. You cannot pin a model you liked, and you will not get a dated notice when it goes. Our guide to AI assistant plans for small teams and the plan comparison weigh that choice.

Bottom line

Open the usage export, find your model ids, and put November 30, 2026, January 6, 2027, February 26, 2027 and April 1, 2027 in a calendar. Then read the retention and data questions in the previous note, because a migration is also a data move.

Sources