Your voice-note workflow has an expiry date you did not set

A template with a date attached to it

The pattern is everywhere: a voice message arrives in Telegram or WhatsApp, the workflow sends the audio to OpenAI, and the text lands in a CRM or a spreadsheet. It is the kind of thing built once, from a template, and then left alone because it works.

It has an end date now, and the person who built it was probably not told.

What OpenAI announced

Four transcription models are going, on one date:

On August 26, 2026, we notified developers using whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe, and gpt-4o-transcribe-diarize of their deprecation and removal from the API on February 26, 2027.

• Deprecations, OpenAI API documentation

The replacements OpenAI names for all four are gpt-transcribe and gpt-live-transcribe, which shipped four weeks before the deprecation was announced.

There is a second, earlier date on the same page that is easy to confuse with this one. A group OpenAI heads as legacy audio, realtime and transcription models shuts down on 20 January 2027, five weeks sooner, and it includes a dated transcription snapshot of its own. If your workflow names a model with a date in it, check which list it is on before you plan around February.

Where the docs say "default" and the code says otherwise

Here is why this lands on people who never chose a model. n8n's OpenAI node has an operation for transcribing audio, and its documentation says:

Use this operation to transcribe audio into text. OpenAI API limits the size of the audio file to 25 MB. OpenAI will use the whisper-1 model by default.

• Audio operations, n8n Docs

The word "default" implies there is a way to choose something else. In the node's source there is not. The operation declares two parameters, the binary field and an options collection holding language and temperature, and then sets the model as a constant before sending the request. Both versions of the node do it, and both files are identical.

That this is deliberate rather than an omission in the documentation is visible on the same page: the sibling operation that generates audio does document a model selector. Transcription has none because none exists.

Which platform lets you choose the model?

Make's field description, verbatim: "Select or map the model you want to use. Refer to the OpenAI Audio API documentation for information on available models. Note: Ensure you have access to the model you want to work with. Only models available for the account are listed for selection." Source: OpenAI modules, Make apps documentation.

The Zapier row is worth reading precisely, because the model is named in one place and absent from another. The action's own field list is file, prompt, response format, timestamp granularity, language and temperature, with no model among them. Zapier's getting-started article for the integration then describes the action itself, verbatim:

Create Transcription - Creates a new transcription, using Whisper, from an audio or video file.

• How to get started with ChatGPT (OpenAI) on Zapier, Zapier Help Center, updated 15 July 2026

So the model is not missing from the interface by accident. It is named in the documentation and not offered as a choice, which puts the Zapier action in the same position as the n8n one: whichever Whisper model it sends today, the date that matters is 26 February 2027, and there is no field to change in advance. Neighbouring actions on the same page do list a model field, so this is a decision about this action rather than a limitation of the integration.

OpenAI's own guide still sends you to the model it is retiring

This is the part that will catch people who go and check. The guide does name a current model as its recommendation, and then tells you to use whisper-1 in three separate places, with no notice that it is going away:

Use whisper-1 when you need word or segment timestamps.

The timestamp_granularities[] parameter is only supported for whisper-1.

To translate a completed audio recording into English, use /v1/audio/translations with whisper-1. Unlike transcription, which preserves the recording's original language, this endpoint returns English text.

• File transcription, OpenAI API documentation

And the model's own page carries no deprecation notice at all: it lists transcription and translation as supported and says nothing about 2027.

The third quote is a genuine open question rather than a stale page. OpenAI documents translation to English as a whisper-1 endpoint and is retiring whisper-1, and nothing found for this article reconciles the two. If your workflow translates rather than transcribes, that is worth asking OpenAI about directly before the date, because the replacement path is not published.

Who actually gets the email?

OpenAI does commit to telling people:

Impacted customers will always be notified by email and in our documentation along with blog posts for larger changes.

• Deprecations, OpenAI API documentation

Read who that is. The notification goes to developers using the model, which means the owner of the API key. In a small business the key is often in a platform account set up two years ago by somebody who has since moved on, and the mailbox it points at may be a shared one nobody reads. The person whose voice notes stop being transcribed is not on that list.

Has anyone asked for a model selector?

Repeatedly, and for about a year. n8n's community forum carries a feature request titled "Update ChatGPT Audio operations node to allow model to be selected" from October 2025, an earlier thread asking the same thing in April 2025, and another in January 2026 noting plainly that there is no model selection in the built-in transcription operation.

What the forum does not carry is anything about the deprecation. Searches for whisper-1 being deprecated, for the replacement model names, and for the 2027 date return nothing. So the gap has been known for a year as an inconvenience, and nobody has yet connected it to a deadline.

The repository says the same: no open issue and no pull request about the hardcoded model, the deprecation, or the migration. The thing that would fix this for everyone is a node update, and the date by which it has to arrive is fixed while the work to make it has not started in public.

What to do before February

1. Find every workflow that transcribes audio. Search for the OpenAI node with an audio resource in n8n, the transcription module in Make, the transcription action in Zapier.
2. Write down which platform each one is on, because that determines whether you have a choice today. On Make you can change the model now. On n8n and on Zapier you cannot.
3. On Make, switch to a named replacement and test it on a real recording, including one in the language your customers actually speak.
4. On n8n, decide between waiting and routing around it. The workaround the community has been posting since 2025 is to call the API from an HTTP Request node, which gives you the model parameter the built-in operation does not expose, at the cost of handling the request yourself.
5. Check whether you translate rather than transcribe. That path is documented only through the model being retired, and it needs its own answer.
6. Find out who receives OpenAI's email for your API key, and make sure it is a mailbox somebody reads.
7. Put 26 February 2027 in the calendar, along with 20 January 2027 if you use realtime audio. The failure will be an API error inside a step, and that is the kind that gets discovered late.

When this is a job to hand over

Changing a model name in a Make module is a five-minute job. Do it yourself.

It becomes a job when the transcription sits in the middle of a workflow that several people depend on, when the same key runs six automations and nobody knows which of them touch audio, or when the replacement changes the shape of the response and everything downstream has to be remapped. A fixed deadline with a countable list of workflows behind it is about the easiest thing there is to quote for, which is what Fix S and Fix M are for, and it is the shape of much of the AI agent work we take.

Another OpenAI deadline with the same shape landed in August, when the Assistants API was switched off. The gap between who was notified and who was affected was the same one.

Sources

• Deprecations - OpenAI API documentation, read 6 September 2026. The four transcription models retiring on 26 February 2027, their named replacements, the separate 20 January 2027 date for legacy audio and realtime models, and the commitment to notify impacted customers by email.
• File transcription - OpenAI API documentation, read 6 September 2026. The recommended model for new work, and the three places the guide still instructs the reader to use whisper-1, including translation to English.
• whisper-1 - OpenAI API documentation, read 6 September 2026. The model page listing transcription and translation as supported, with no deprecation notice.
• Audio operations - n8n Docs, read 6 September 2026. The transcription operation's documented parameters, the statement that OpenAI will use whisper-1 by default, and the sibling audio-generation operation that does document a model selector.
• n8n-io/n8n - n8n source, read 6 September 2026. The transcription operation setting the model as a constant, the absence of a model property in both node versions, and the absence of any open issue or pull request about the deprecation.
• Update ChatGPT Audio operations node to allow model to be selected - n8n Community, 17 October 2025, read 6 September 2026. A feature request for the model selector the transcription operation does not have, still open.
• OpenAI modules - Make apps documentation, read 6 September 2026. The transcription module's Model field and its note that only models available to the account are listed.
• How to get started with ChatGPT (OpenAI) on Zapier - Zapier Help Center, updated 15 July 2026, read 7 September 2026. The action list describing Create Transcription as using Whisper, and the neighbouring actions that do name a model field.