Transcription Model Retirement Dates for Captioning: Azure October 15, OpenAI February 26

The first transcription model retirement for captioning teams lands on October 15, 2026. That day, Azure OpenAI retires the original 2025-03-20 versions of gpt-4o-transcribe and gpt-4o-mini-transcribe, along with gpt-4o-mini-tts and a sora-2 version. Azure’s whisper, tts and tts-hd follow on December 15. On OpenAI’s own API, whisper-1 and the whole gpt-4o-transcribe family shut down on February 26, 2027.

This matters to associations and universities because transcription models sit behind a lot of captioning work. Webinar recordings, conference sessions and lecture captures often reach a model through a WordPress plugin, a Zapier step or a short script someone wrote two years ago. This piece lists each date by platform with the replacement the vendor names, shows how to find which model your tools call, and covers what to test before you switch. Captions matter for access, and a silent failure means recordings go out without them.

The short version: check Azure deployments this week, because Azure lists no replacement for most of these rows. On OpenAI, plan the move to gpt-transcribe or gpt-live-transcribe well before February. Do not assume the new models do everything the old ones did. OpenAI’s own documentation, as of late September 2026, ties word timestamps to whisper-1, and caption files depend on timing.

Transcription model retirement dates on Azure OpenAI

All Azure dates below come from Microsoft’s model retirement schedule, last updated September 21, 2026, and read on September 27. The replacement column on that page is blank for every audio row.

Two details in that table change the plan. First, the October 15 transcription model retirement covers the 2025-03-20 versions only. A newer gpt-4o-mini-transcribe version, dated 2025-12-15, stays on the schedule until June 15, 2027. Second, gpt-4o-transcribe has no newer version on the schedule at all. If a caption workflow depends on it, the next step is a different model.

What happens on Azure after a transcription model retirement

Microsoft’s lifecycle and support policy says a retired model is removed from service and inference requests return a 410 Gone error. A caption job that calls a retired deployment will fail rather than fall back.

The same page describes automatic upgrades for Standard, Global Standard and Data Zone Standard deployments, rolled out region by region. Provisioned deployments are not upgraded and need a manual move. That policy is written mostly with chat models in mind. With no replacement listed for these audio rows, we could not confirm what an automatic upgrade would target. Treat any transcription deployment as needing a manual check.

Azure does offer newer options. The Foundry models list, updated September 23, includes gpt-transcribe for file transcription through /audio/transcriptions. It notes that gpt-transcribe is not a Realtime API model. For streaming, it points to gpt-realtime-whisper or gpt-live-transcribe. Region availability sits on a separate page, so confirm your region before planning a move.

One boundary is worth stating. Azure AI Speech is a separate service with its own batch transcription, including a Whisper option. Its Whisper overview page, updated August 5, mentions no retirement date. This schedule covers Azure OpenAI model deployments in Foundry.

OpenAI’s transcription model retirement dates: January 20 and February 26

OpenAI’s deprecations page carries two entries that touch transcription. Each names replacements, which Azure’s schedule does not.

The January entry, announced July 20, 2026, is mostly about legacy realtime and audio families. gpt-realtime and gpt-4o-realtime move to gpt-realtime-2.1, and gpt-audio and gpt-4o-audio move to gpt-audio-1.5. Live caption tools built on those Realtime API models fall under that date.

The February entry, announced August 26, is the one most captioning scripts will hit. OpenAI’s page says it notified developers using those four models. Check the inbox tied to your API account for that notice.

The snapshot trap in OpenAI’s dates

The January row tells users of the March 2025 snapshot to move to the December 2025 snapshot. The February row then lists gpt-4o-mini-transcribe without a snapshot date. We read that as the whole family, but OpenAI does not spell it out. Either way, the December snapshot is a short stop, not a destination. Moving there in January buys about five weeks.

What OpenAI has not listed

We found no entry for tts-1, tts-1-hd or gpt-4o-mini-tts on OpenAI’s deprecations page as of September 27. If your audio work includes spoken versions of articles, OpenAI’s text-to-speech models are not on this transcription model retirement list. Azure’s text-to-speech rows are.

The October 23 list is a separate shutdown

OpenAI has a second, unrelated shutdown on October 23, 2026. That list has 17 rows, including gpt-4-turbo, o1, o3-mini, gpt-image-1 and several fine-tuned bases. No audio, realtime or transcription model appears on it.

So a team that cleared its October 23 work has not cleared the transcription model retirement work. The reverse is also true. Treat them as two projects with two owners if needed. The pattern is familiar from the Assistants API shutdown in August: the vendor publishes the date, and the work lands on whoever maintains the integration.

How to find which transcription model your captioning tool uses

Most captioning setups hide the model name. Here is where to look, in rough order of effort.

• Plugin and platform settings: open the settings page of every caption, transcript or media plugin. Many show the model as a dropdown or a text field. If a setting says only “Whisper,” ask the developer which API and model ID it calls.
• Zapier steps: Zapier’s OpenAI integration page describes its Create Transcription action as using Whisper. The page does not show a model picker for it. Zapier controls which model that action calls, so watch its changelog and test a run after February.
• Azure deployments: in the Foundry portal, each deployment shows its model name and version. Anything on `gpt-4o-transcribe` or `gpt-4o-mini-transcribe` version 2025-03-20 is on the October 15 list.
• Custom scripts and repositories: search code, environment files and scheduled jobs for the model strings.
• OpenAI notices: the deprecation email names the models your account has used.

A text search catches most hardcoded model IDs. Run something like this from the root of a project or a server’s web directory:

grep -rnE "whisper-1|gpt-4o-(mini-)?transcribe|gpt-4o-transcribe-diarize|gpt-4o-mini-tts|\"whisper\"|tts-hd" .

In WordPress, check the wp_options table too, since plugins often store the model choice there. We used the same method for the Gemini shutdown dates earlier this fall.

What the replacements document, and what they leave out

This is where the transcription model retirement gets harder than a string swap. As of September 27, OpenAI’s documentation points some caption features at the models that are going away.

OpenAI’s speech to text guide recommends gpt-transcribe for general file work. It also says to use whisper-1 when you need word or segment timestamps. It states that the timestamp_granularities[] parameter is supported only for whisper-1. Speaker labels come from gpt-4o-transcribe-diarize.

The realtime transcription guide says gpt-live-transcribe does not return word-level timestamps, speaker labels or confidence scores. It suggests a file transcription model or an application fallback when you need them.

OpenAI may add these features before February, and its guides change often. For now, plan as if timing, speakers and caption files could need a second step, such as a separate alignment tool.

What to test before you switch

Run the same small set of real recordings through the old model and the new one. Pick a webinar with a guest speaker, a lecture with technical terms, and a panel with cross-talk. Then compare these points.

• Names and jargon: `gpt-transcribe` accepts `prompt`, `keywords` and `languages` hints, according to OpenAI’s guide. Feed it speaker names, program names and acronyms. OpenAI notes that keywords are hints, not guaranteed output.
• Timestamps: confirm the new model returns timing you can build captions from. Check that lines stay in sync late in a long recording, not only in the first minutes.
• Output format: request SRT and VTT directly. If the model returns only text or JSON, budget time to build caption files another way.
• Speakers: if panels need speaker labels, test that path on its own. The diarization model shuts down on OpenAI in February.
• Languages: test every language you caption. Language hints use ISO codes, and accuracy varies by language.
• Long files: OpenAI’s guide and Azure’s model list both give a 25 MB limit per request. Check how your tool splits longer lecture captures.
• Live versus recorded: live webinar captions and post-event captions may need different models, `gpt-live-transcribe` for the first and `gpt-transcribe` for the second.

Keep the human review step. A model swap can change error patterns in ways a quick skim misses. Automated checkers, such as the engine in our Deque axe explainer, can inspect page markup. They cannot tell you whether the words in a caption track are right.

Price per minute, where OpenAI documents it

OpenAI publishes per-minute prices for three of the models involved. We have not checked Azure pricing for this piece.

The gpt-4o-transcribe models are billed by audio tokens rather than by the minute. Prices were read from OpenAI’s model pages on September 27 and can change. Your caption tool may also add its own markup on top.

A dated plan for captioning teams

Put each transcription model retirement date on the calendar of whoever owns captions, not only the developer.

• Now to October 15: find every Azure deployment on a 2025-03-20 transcription or speech version and move it. Retired calls return errors.
• By December 15: move off Azure `whisper`, `tts` and `tts-hd`.
• By January 20: move off OpenAI’s legacy realtime and audio families and the March 2025 mini transcribe snapshot.
• By February 26: move every OpenAI `whisper-1` and `gpt-4o-transcribe` family call, after testing timing and caption output.

The spring semester and conference season both start before late February. Testing in November leaves room to fix caption output before those recordings pile up.

Frequently Asked Questions