The dropdown still says GPT-5, and that is not reassurance

What is actually retiring

Six dated snapshots, on one day:

On June 11, 2026, we notified developers using older GPT-5 and o3 model snapshots of their deprecation and removal from the API on December 11, 2026.

• Deprecations, OpenAI API documentation

Those replacements are what the page says today, and they have changed at least once: a developer quoting the same table in June reproduced different model names in those rows. Take them from the page when you do the work, not from an article, this one included.

"But my workflow just says gpt-5"

It feels like safety and it is not. OpenAI's model page for GPT-5 lists the snapshots the plain name covers, and there is exactly one: the dated snapshot that is retiring. The same holds for the mini, nano and pro variants, and for both o3 entries.

So a workflow set to the undated name is, today, sending requests to the snapshot on that table. The plain name is not an alternative to the dated one. It is a pointer at it.

What OpenAI has not published is what happens to that pointer on 12 December: whether the plain name re-points to something current, or stops resolving along with its snapshot. A developer asked precisely that five days after the announcement, in June. Another user replied; OpenAI did not, and the question is still open.

Where a snapshot gets pinned in your tools

This lands on automations harder than on developers, because platforms store the choice you made and go on storing it after the dropdown has moved on.

n8n offers a list or a typed id. The node's source describes the field as "Choose from the list, or specify an ID", loads the list from the models available to your account, and ships with a default of gpt-5-mini. So the value stored in your workflow is a string somebody picked, possibly two versions of the node ago.

Make documents the same shape: "Select or map the model you want to use", with the note that only models available to your account are listed. A mapped value is worse for auditing than a chosen one, because it may be assembled at runtime and never appear in the module at all.

Zapier describes its ChatGPT actions in terms of model families, without dated ids, in the pages published for them. Whether the editor lets you type an arbitrary id is not something those pages settle, so open the action and look.

Make has already documented what its dropdown does

This is the most useful evidence available, because it is not a prediction. Make has published notes on three separate OpenAI retirements, and it behaved differently each time.

In March 2026, a silent reroute, in Make's own words:

Existing scenarios that use deprecated models might still display the model name in the model dropdown, but will use gpt-4.1 for future requests. [...] Outputs may vary slightly due to differences in model behavior.

• OpenAI model deprecations on March 26, 2026, Make Help Center

Read that twice, hedges included. The interface might still show the model you chose, the requests go to a different one, and the vendor's phrasing for the consequence is that outputs may vary slightly.

In February 2026, a silent substitution, with the scenario quietly rewritten:

We've removed the deprecated chatgpt-4o-latest model from the OpenAI app and automatically migrated existing scenarios to gpt-4o-2024-08-06. No action is required on your part.

• Deprecated OpenAI model automatically replaced, Make Help Center

And in August 2026, the opposite promise, for a different set of models:

After this date, scenarios using these deprecated models will stop running.

• OpenAI: Deprecation of select GPT models by August 10, 2026, Make Help Center

Same vendor, same year, three different outcomes: rerouted, substituted, stopped. Make announced each of them beforehand, so December's answer will probably appear on its help centre in November - which is exactly the page to check, because "will my scenario break" has more than one possible answer. It might break, or it might quietly start answering your customers with a different model.

What the failure looks like when it is a hard stop

OpenAI does not document an error code, status or fallback for calling a retired model. What exists is a field report: a Zapier user in 2024, whose Zap returned a message saying the model had been deprecated, with the model name in it.

Take that as the shape of the failure, not as a specification. If a step in your automation starts failing in December with a message naming a model, this is what happened, and the fix is a string.

The other dates in the same window

Worth putting in the same calendar entry, because they land within weeks of each other and each has its own audience:

All from the same deprecations page, which is the single page worth bookmarking out of everything in this article.

Also worth knowing, from OpenAI's stated policy: generally available models get at least six months of notice, specialised variants at least three, and preview models "may be retired with much shorter notice, such as 2 weeks". If your workflow uses anything with "preview" in the name, that is the shortest fuse on the list.

What to do before December

1. List every place a model name is stored. Every AI node in n8n, every OpenAI module in Make, every ChatGPT step in Zapier, plus anything calling the API from a code step or an HTTP request.
2. Write down the exact string in each, dated or not. "GPT-5" is not an answer; gpt-5 and gpt-5-2025-08-07 are, and today they point at the same place.
3. Check the mapped ones especially. A model chosen from a list is visible. A model assembled from a variable is not, and it will not be found by looking at the module.
4. Move to a current model in a copy of the workflow first, and compare outputs on real inputs. Model changes alter behaviour, which is why Make's own note about rerouting mentions varying outputs.
5. Do not assume your platform will migrate you. Make has done all three things in one year: rerouted, substituted and stopped.
6. Set a reminder for the last week of November, which covers the prompts and Agent Builder shutdowns as well as leaving room before the snapshot date.
7. Watch the runs after the date. If your platform reroutes silently, nothing errors and the only signal is that answers changed, which is the hardest kind of failure to notice.

When this is a job to hand over

Changing a model name in three workflows is an afternoon. Do it yourself.

It becomes a job when nobody can list where the model names live, when the automation's output feeds something that has to keep behaving the same way, or when the model change turns out to alter formatting that six downstream steps depend on. A model swap you can describe in advance is a job you can price in advance, which is what Fix S and Fix M exist for, and the AI agent work we take is largely made of them.

Three OpenAI dates this year arrive through a platform instead of through a developer. The Assistants API shutdown in August was the loudest, the transcription models go in February, and this one sits between them. The pattern behind all three is that the notice is written for whoever built the thing, not for whoever depends on it.

Sources

• Deprecations - OpenAI API documentation, read 6 September 2026. The 11 December 2026 shutdown of six GPT-5 and o3 snapshots with their recommended replacements, the adjacent October, November, December and January dates, and the notice periods for generally available, specialised and preview models.
• GPT-5 model page - OpenAI API documentation, read 6 September 2026. The default snapshot for the plain model name, and the list of snapshots and aliases it covers. The mini, nano, pro and o3 pages carry the same structure.
• OpenAI model deprecations on March 26, 2026 - Make Help Center, read 6 September 2026. Scenarios continuing to display the old model name in the dropdown while sending requests to a different model, and the note that outputs may vary.
• Deprecated OpenAI model automatically replaced - Make Help Center, read 6 September 2026. A deprecated model removed from the app with existing scenarios migrated automatically.
• OpenAI: Deprecation of select GPT models by August 10, 2026 - Make Help Center, read 6 September 2026. The note that scenarios using those deprecated models will stop running after the date.
• n8n-io/n8n - n8n source, read 6 September 2026. The OpenAI chat model node describing its field as a list or a specified id, loading the list from the account, and defaulting to gpt-5-mini.
• Deprecation of GPT-5 and o3 snapshots - OpenAI Developer Community, 16 June 2026, read 6 September 2026. A developer asking what happens to the undated model names after the shutdown, answered by another user and not by OpenAI, and quoting the replacement names as the table listed them in June.
• OpenAI modules - Make apps documentation, read 6 September 2026. The model field description, including selecting or mapping the model and the note about models available to the account.