OpenAI Shuts Off Seventeen Models on October 23, and the Fine-Tuned Ones Do Not Come Back

The largest of this year’s OpenAI model deprecations lands on October 23, 2026, when seventeen model identifiers stop working in the API. Six of them are fine-tuned models, and the page announcing it also contains a sentence saying fine-tuned models are not affected.

Both statements are on the same page. Only one of them applies. Anyone skimming for reassurance will find the wrong one.

We covered the Assistants API shutdown in August, when twelve months of notice ran out on thousands of production apps. This is larger, it is sixteen days away, and the part that cannot be undone is the fine-tuning.

What shuts off on October 23

The deprecations page announced this wave on April 22, 2026 under the heading "Legacy GPT model snapshots." Every identifier below has a shutdown date of 2026-10-23.

Base models: gpt-3.5-turbo-0125, gpt-4-0613, gpt-4-1106-preview, gpt-4-turbo, gpt-4.1-nano, gpt-4o-2024-05-13, gpt-image-1, o1-2024-12-17, o1-pro-2025-03-19, o3-mini-2025-01-31, o4-mini-2025-04-16.

Fine-tuned models, listed as their own rows with their own shutdown date: ft-gpt-3.5-turbo, ft-gpt-4, ft-gpt-4.1-nano-2025-04-14, ft-o4-mini-2025-04-16, ft-babbage-002, ft-davinci-002.

Two entries on that list deserve pointing at, because they are not what a reader scanning for "GPT-4 and the o-series" would expect to find.

gpt-4o-2024-05-13 is a GPT-4o snapshot, not a GPT-4 one. If you pinned that snapshot in 2024 for reproducibility, which was the responsible thing to do at the time, you are on this list.

gpt-4.1-nano is on it too, and it is listed as a bare alias rather than a dated snapshot, along with gpt-4-turbo and gpt-image-1. Everything else is snapshot pinned.

Worth getting the vocabulary right, because OpenAI’s own definitions carry the whole meaning. From the page: "When we announce that a model or endpoint is being deprecated, it immediately becomes deprecated. All deprecated models and endpoints will also have a shut down date. At the time of the shut down, the model or endpoint will no longer be accessible." So every model here has been deprecated since April. October 23 is when it is gone.

Three smaller waves have already passed while most people were watching this one. On September 24 the Videos API and the entire Sora 2 family shut down. On September 28, gpt-3.5-turbo-instruct, babbage-002, davinci-002 and gpt-3.5-turbo-1106 went. On October 1, gpt-5.4-cyber went.

The sentence that says the opposite of what happens

This is the part worth reading twice, and it is the reason this post exists.

Search the deprecations page for "fine-tun" and you will find this: "Fine-tuned models created from these base models are not affected by this deprecation, but you will no longer be able to create new fine-tuned versions with these models."

That sentence appears twice. Both times it sits under historical entries from 2023, one dated 2023-08-22 and one dated 2023-11-06. It governs nothing in 2026. And it is the single most likely thing on the page to mislead somebody, because it is reassuring, it is about fine-tuning, and it is on the same page as the thing that contradicts it.

The 2026 statement is in a section headed "Update to OpenAI’s self-serve fine-tuning," announced May 7, 2026: "Inference on fine-tuned models will be disabled only when the underlying base model is deprecated."

That sentence is also wrong, in a different way. By the page’s own glossary, deprecation is the announcement. These bases were deprecated on April 22, 2026. Fine-tuned models built on them are still serving requests today, five months later. Read strictly against the definition printed further up the same page, the sentence is already false.

The reliable signal is neither sentence. It is the table. Six ft- rows carry a shutdown date of 2026-10-23, and shutdown is the word OpenAI defines as "no longer be accessible." Read "deprecated" in that prose as "shut down" and the picture resolves.

The door closes twice, and the second time is permanent

Here is the consequence that most openai model deprecations coverage has missed, because it requires reading two separate sections against each other.

The self-serve fine-tuning wind-down has its own schedule, quoted verbatim:

• May 7, 2026: creating fine-tuning jobs is no longer available to organizations that have not previously run fine-tuning.
• July 2, 2026: creating fine-tuning jobs is no longer available to organizations that have not run inference on a fine-tuned model in the past 60 days.
• January 6, 2027: "Active existing customers will no longer be able to create new fine-tuning jobs on this date."

Put that beside October 23 and the window is visible. A fine-tuned model in production stops answering on October 23. The ability to create a new fine-tuning job, on any base, ends on January 6, 2027.

That is roughly ten weeks to retrain onto a supported base, and only if you qualify as an active existing customer. After January 6 the option is not delayed or degraded. It is closed.

If you have a fine-tuned model doing real work, that is the deadline on this page that actually constrains you, and it is not the one in the headline.

There is a second wave seven weeks later

October 23 is not the end of it, and the follow-on catches a different group of people: the ones who already migrated once.

Announced June 11, 2026, with a shutdown date of December 11, 2026: gpt-5-2025-08-07, gpt-5-mini-2025-08-07, gpt-5-nano-2025-08-07, gpt-5-pro-2025-10-06, o3-2025-04-16 and o3-pro-2025-06-10.

Anyone who moved off GPT-4 onto GPT-5 in 2025, which was the advice at the time, has a second migration landing seven weeks after the first.

Two other December items are easy to miss. On December 1, gpt-image-1-mini, gpt-image-1.5 and chatgpt-image-latest all shut down, with gpt-image-2 as the replacement. Note the ordering: gpt-image-1 dies on October 23 and the newer image models die six weeks later. Migrate gpt-image-1 to anything other than gpt-image-2 and you will migrate twice.

And on November 30, three things that are not models at all go away: the v1/prompts API and reusable prompt objects, the Evals dashboard and API, and Agent Builder. The recommended replacement for Evals is Promptfoo, which is a third-party tool. If any of your workflow lives in OpenAI’s own evaluation tooling, that is a November problem with a non-OpenAI answer.

One entry worth flagging for what it says about notice rather than what it shuts down. gpt-5.4-cyber was announced on September 11 and shut down on October 1. Twenty days. OpenAI’s published policy is "at least 6 months" for generally available models and "at least 3 months" for specialized variants, with short notice reserved for models carrying preview in the name. gpt-5.4-cyber is not a preview model. The policy carves out safety and compliance reasons, which a cyber model plausibly falls under, and no justification is given in the entry.

The page and the tooling point at different destinations

A smaller inconsistency, and one that could waste real engineering time if nobody notices it.

The deprecations page recommends gpt-5.6-sol, gpt-5.6-terra and gpt-5.6-luna as the replacements for almost everything on the October list. We wrote about the GPT-5.6 family when it shipped.

But OpenAI’s own guidance page on choosing a current model is written around migrating to GPT-6 Astra, and the only migration automation OpenAI ships, the Codex "OpenAI Docs" skill invoked as a migration command, targets Astra rather than 5.6.

So the table says one thing and the tooling assumes another. Neither is wrong exactly. They are answering different questions: the table is answering "what is the nearest equivalent," and the guidance is answering "what should you build on now." Worth deciding which question you are asking before committing a sprint to it, because the two answers have different prompt behavior and different pricing.

Since that table was written, a third option appeared. OpenAI released GPT-6 Sol and Luna on September 22 at $2.00 input and $10.00 output for Sol, which is half what GPT-5.6 Sol costs. So the destination the deprecations page names is now twice the price of a newer model in the same position, and the page has not been updated to say so. If you are migrating before October 23, landing on GPT-6 rather than on the GPT-5.6 models the table names saves half the running cost and avoids a second migration later.

There is no migration guide, script, or tool specific to the October 23 wave. No prompt translation, no eval harness for moving GPT-4 or o1 prompts onto a successor, no aliasing of retired identifiers to replacements. OpenAI has published migration material for Agent Builder, for Evals and for the Assistants API, and nothing for this.

Sora is the one with no replacement at all

Every other entry on the deprecations page in this window names a recommended replacement. One does not.

The September 24 shutdown covered the Videos API itself plus sora-2, sora-2-pro, sora-2-2025-10-06, sora-2-2025-12-08 and sora-2-pro-2025-10-06. The recommended replacement column is empty for all six rows, including for the API endpoint. There is no successor video model or endpoint listed anywhere on the page.

That is not an oversight in a table. In the same window OpenAI named replacements for eleven separate shutdowns, some of them for models nobody uses. Leaving the column blank across an entire product surface reads as an exit, and it deserves more attention than it got.

Nothing here touches ChatGPT

Worth saying plainly, because this is where client conversations go sideways.

The deprecations page does not mention ChatGPT at all. It opens by scoping itself: "This page lists all API deprecations, along with recommended replacements." Everything above is an API and integration problem.

ChatGPT retirements run on a separate track with different dates, and they already happened. GPT-4o, GPT-4.1, GPT-4.1 mini and o4-mini were retired from ChatGPT on February 13, 2026, and the GPT-5.1 family on March 11, 2026. The announcement said, verbatim, "In the API, there are no changes at this time."

So the two tracks ran in opposite order. ChatGPT dropped o4-mini in February while the API kept it until October. If a client asks whether the October date affects their team’s ChatGPT seats, the answer is no.

Where the model ID is actually hiding

For most of our readers the affected identifier is not in a repository at all, which is why a code search comes back clean and the integration breaks anyway.

Check the database before the codebase. WordPress AI plugins store the configured model in the options table. Drupal’s AI module and its provider submodules store it in config YAML, which may or may not be exported to the repository depending on how the site was built. A site built two years ago by somebody else will have a model identifier sitting in configuration that nobody has looked at since launch.

Check anything that wraps OpenAI rather than calling it. Content generation plugins, chat widgets, semantic search add-ons, alt-text generators and translation modules all pin a model somewhere, and their vendors are on their own release schedules. A plugin that has not shipped an update since 2024 is pinned to something on this list with reasonable odds.

Check environment variables and secrets stores, not just committed files. A serverless function or a worker with OPENAI_MODEL set in a hosting dashboard will not appear in any grep.

Check saved prompt objects specifically. Those disappear on November 30 along with the v1/prompts API, separately from the model shutoff, and they live on OpenAI’s side rather than yours.

Then check the logs. The most reliable audit is not a search at all: pull the model field from your own request logging for the last thirty days and count the distinct values. That catches everything above at once, including the integrations you had forgotten about, and it is the only method that finds a model identifier set by a person rather than by a file.

What to do before October 23

Grep for model identifiers, not model names. The list is snapshot pinned. A codebase that says gpt-4o-2024-05-13 is affected and a codebase that says gpt-4o is not, and no amount of thinking about "are we on GPT-4" will surface the difference. Search for the literal strings.

Find your fine-tunes first, and decide this month. They stop on October 23, and the ability to create a replacement ends on January 6, 2027. That is the only item here with a door that closes for good.

Check whether you are in the December wave too. If you migrated to GPT-5 in 2025, gpt-5-2025-08-07 and its siblings go on December 11. Doing both migrations as one piece of work is cheaper than doing them seven weeks apart.

If you use gpt-image-1, go to gpt-image-2 and nowhere else. The intermediate image models die on December 1.

Move your Evals work before November 30, and know that OpenAI’s recommended destination is a third-party tool.

Pick a destination deliberately. The table says GPT-5.6. The current-model guidance and the migration automation say GPT-6 Astra. Choose on your own criteria rather than following whichever page you happened to open.

Do not build alerting on the error you expect to get. We could not verify what OpenAI’s API returns for a shut-down model from any OpenAI documentation. The error codes guide documents 400, 401, 403, 429, 500 and 503 and contains no case for an unknown or retired model. Community reports consistently show a 404 with model_not_found, but that is user reporting, not documentation, and the same error covers "does not exist" and "you do not have access," so it is indistinguishable from an entitlement problem. Alert on a drop in successful completions instead.

Sixteen days is enough time for the grep. It is not much time for the fine-tunes.

Frequently Asked Questions