SunsetDiff

Azure OpenAI gpt-4o-mini-tts, transcribe, and realtime audio retirement on June 15, 2027

The dataset has 4 Azure OpenAI ids with shutdown date 2027-06-15: gpt-4o-mini-transcribe-2025-12-15, gpt-4o-mini-tts-2025-12-15, gpt-audio-mini-2025-12-15, gpt-realtime-mini-2025-12-15. The retirement schedule lists that date on each of those rows. Lifecycle is GA. The replacement cell is empty, so those cells say Not published. The lifecycle page says a Retired model is removed from service and all inference requests return 410 Gone. gpt-4o-mini-transcribe is file speech-to-text. gpt-4o-mini-tts is text-to-speech. gpt-audio-mini is chat audio. gpt-realtime-mini is the Realtime API. A voice app has to retest the call it already sends.

Last reviewed . Publisher: SunsetDiff.

Steps

  1. Find deployments. Call Deployments List, or Get-AzCognitiveServicesAccountDeployment, for each Azure OpenAI resource. Read properties.model.name and properties.model.version. A deployment name can differ from the model name. gpt-4o-mini-transcribe version 2025-12-15 is the dataset id gpt-4o-mini-transcribe-2025-12-15. gpt-4o-mini-tts version 2025-12-15 is gpt-4o-mini-tts-2025-12-15. gpt-audio-mini version 2025-12-15 is gpt-audio-mini-2025-12-15. gpt-realtime-mini version 2025-12-15 is gpt-realtime-mini-2025-12-15.
  2. Pick a replacement. The replacement cell is empty for all 4 rows. Microsoft did not name a recommended replacement. Shortlist a model yourself. gpt-audio-1.5 version 2026-02-23, gpt-realtime-1.5 version 2026-02-23, and gpt-5.3-codex version 2026-02-24 retire on 2027-08-24. The June 15 replacement cells do not name those ids.
  3. Test. Send one real audio request to the model you picked, using a clip or prompt you send now. Do that before June 15, 2027. Check latency, the voice, the audio format, and the API version on that call.
  4. Redeploy before the date. Create a deployment of the model you tested and move traffic before June 15, 2027. These rows do not name an automatic-upgrade target. Provisioned deployments are not auto-upgraded. A deployment set to NoAutoUpgrade stops working on the retirement date.

Models and dates

These 4 ids are the Azure OpenAI rows whose shutdown date is 2027-06-15. The model id links to its SunsetDiff model page. Lifecycle is the schedule's Lifecycle column, copied from the source excerpt stored on the row. Shutdown date and replacement are the fields stored on the row. Not published means the replacement field is empty. These rows do not store a deprecation_date. gpt-4o-mini-transcribe-2025-12-15 (GA), gpt-4o-mini-tts-2025-12-15 (GA), gpt-audio-mini-2025-12-15 (GA), gpt-realtime-mini-2025-12-15 (GA).

Checked on 2026-11-05, the retirement schedule lists gpt-4o-mini-transcribe version 2025-12-15, gpt-4o-mini-tts version 2025-12-15, gpt-audio-mini version 2025-12-15, and gpt-realtime-mini version 2025-12-15 as GA with retirement date 2027-06-15 and an empty replacement cell. Those rows do not carry footnote 2 or footnote 3.

The same schedule prints gpt-realtime-mini version 2025-12-15 a second time. That later row says GA and retirement 2026-12-15, and its replacement cell is empty. SunsetDiff keeps the first base-table row when the same model and version is repeated. The stored shutdown date stays 2027-06-15. The November and December page records the same rule: Azure OpenAI retirements from November 15 to December 16, 2026.

The lifecycle table says Generally Available means: "Production-ready. Weights and APIs are fixed." New deployments are allowed, and existing deployments work, until the retirement date. The lifecycle page also says: "Check the Model retirement schedule for the current retirement dates and replacement model, since these details are subject to change."

Other versions of these names are a different shutdown date. gpt-4o-mini-transcribe-2025-03-20 stores shutdown date 2026-10-15. gpt-4o-mini-tts-2025-03-20 stores shutdown date 2026-10-15. gpt-audio-mini-2025-10-06 stores shutdown date 2027-04-06. gpt-realtime-mini-2025-10-06 stores shutdown date 2027-04-06. They are not in the table on this page.

The Azure Government schedule does not list gpt-4o-mini-transcribe, gpt-4o-mini-tts, gpt-audio-mini, or gpt-realtime-mini. The dates in the first table are the public-cloud schedule.

4 Azure OpenAI model ids with shutdown date 2027-06-15.
Model idLifecycle on the scheduleShutdown dateRecommended replacement
gpt-4o-mini-transcribe-2025-12-15GA2027-06-15Not published
gpt-4o-mini-tts-2025-12-15GA2027-06-15Not published
gpt-audio-mini-2025-12-15GA2027-06-15Not published
gpt-realtime-mini-2025-12-15GA2027-06-15Not published

Sources: Azure OpenAI retirement schedule , Foundry Models lifecycle , Azure Government retirement schedule , checked November 5, 2026.

Which call each model uses

The architecture page says Azure OpenAI provides audio models through three interfaces: the Realtime API for low-latency voice conversations, the Chat Completions API with audio for flexible audio generation and transcription in a single model call, and the Audio API via the /audio endpoint for file-based transcription, translation, and TTS.

gpt-4o-mini-transcribe is on Audio API /audio/transcriptions. The architecture table says: "File-based transcription of prerecorded audio". The transcription reference is POST /openai/v1/audio/transcriptions. Its model options include gpt-4o-mini-transcribe-2025-12-15. That page does not print a date-based api-version.

gpt-4o-mini-tts is on Audio API /audio/speech. The architecture table says: "The tts-hd model is optimized for quality, and gpt-4o-mini-tts supports prompt-guided style and tone." The audio generation quickstart lists gpt-4o-mini-tts as fast, scalable text-to-speech.

gpt-audio-mini is on Chat Completions. The architecture table says: "Combines audio input and output with language reasoning, summarization, or generation in a single model call". The audio generation quickstart says: "Modalities supported in gpt-audio and gpt-audio-mini models include: text, audio, and text + audio."

gpt-realtime-mini is on the Realtime API. The architecture table says: "Low-latency, speech-in and speech-out conversations for live voice agents, interactive assistants, and streaming audio scenarios". The realtime how-to lists gpt-realtime-mini version 2025-12-15 among its supported models. It says: "For all Realtime API models, use the GA endpoint format with /openai/v1 in the URL." And: "For the Realtime API, use the GA endpoint with /openai/v1 in the URL. Don't use date-based API versions or the api-version query parameter."

A speech-to-text app, a text-to-speech app, a chat audio app, and a realtime app are four calls. The schedule does not name a model that keeps the same call. When you shortlist a target, confirm that target still accepts the call you send now.

Sources: Azure speech recognition and generation , Azure OpenAI audio generation quickstart , GPT Realtime API , Create a transcription , checked November 5, 2026.

Replacements Microsoft names

The schedule says: "The Replacement column identifies Microsoft's recommended replacement and, where supported, the target for automatic upgrades at retirement." It also says: "You don't need to wait for a replacement to be listed before evaluating or manually migrating to another compatible model." The lifecycle page says: "The replacement listed in the Model retirement schedule is Microsoft's recommended migration target and, where supported, the model used for automatic upgrades at retirement. It doesn't limit your manual migration choices."

No row in this table stores a replacement.

The replacement cell is empty on gpt-4o-mini-transcribe-2025-12-15, gpt-4o-mini-tts-2025-12-15, gpt-audio-mini-2025-12-15, gpt-realtime-mini-2025-12-15. Empty means Microsoft did not name a recommended replacement on that row. This guide does not add one.

The lifecycle page says: "When you choose a target, evaluate candidate models against your own application, prompts, and representative data. Compare quality, latency, and cost together rather than relying on public benchmarks alone." Then: "Before you migrate, confirm API compatibility, feature support, deployment-type and regional availability, capacity, and quota."

Sources: Azure OpenAI retirement schedule , Foundry Models lifecycle , checked November 5, 2026.

What to test

Send one real request before June 15, 2027, using audio you send now. The lifecycle page says: "When you choose a target, evaluate candidate models against your own application, prompts, and representative data. Compare quality, latency, and cost together rather than relying on public benchmarks alone." That sentence names latency. These pages do not print one latency number that covers all 4 models.

The realtime how-to says: "In most cases, we recommend using the WebRTC API for low-latency real-time audio streaming." Its connection-methods table lists WebRTC at ~100ms, WebSocket at ~200ms, and SIP as Varies. The transcription reference says: "Supplying the input language in ISO-639-1 (e.g. en) format will improve accuracy and latency." It also says: "There is no additional latency for segment timestamps, but generating word timestamps incurs additional latency." The audio generation quickstart does not print a latency number for gpt-4o-mini-tts or gpt-audio-mini.

Voices: the audio generation quickstart says: "The following voices are supported for audio out: Alloy, Ash, Ballad, Coral, Echo, Sage, Shimmer, Verse, Marin, and Cedar." The supported-models table on that page includes gpt-4o-mini-tts, gpt-audio-mini, and gpt-realtime-mini. It does not include gpt-4o-mini-transcribe. A transcription call returns text. The realtime how-to shows an example session with voice alloy and input_audio_format pcm16. Test the voice string your app sends.

Audio formats: the same quickstart says: "The following audio output formats are supported: wav, mp3, flac, opus, pcm16, and aac." And: "The maximum audio file size is 20 MB." The realtime how-to says: "The Realtime API expects audio in a specific format:" The list under that sentence is PCM 16-bit (pcm16), mono (single channel), and 24kHz. The transcription reference says: "The audio file object (not file name) to transcribe, in one of these formats: flac, mp3, mp4, mpeg, mpga, m4a, ogg, wav, or webm." It also says: "For gpt-4o-transcribe and gpt-4o-mini-transcribe, the only supported format is json."

API version: the audio generation quickstart says: "Support for audio completions was first added in API version 2025-01-01-preview." The sample on that page sets the model to gpt-audio-mini. The realtime how-to says: "For the Realtime API, use the GA endpoint with /openai/v1 in the URL. Don't use date-based API versions or the api-version query parameter." The transcription reference uses POST /openai/v1/audio/transcriptions and does not print a date-based api-version. The schedule does not print an api-version for gpt-4o-mini-tts version 2025-12-15. Keep the api-version the speech call already sends, and send that same version to the model you picked.

Sources: Azure speech recognition and generation , Azure OpenAI audio generation quickstart , GPT Realtime API , Create a transcription , Foundry Models lifecycle , checked November 5, 2026.

What happens to a deployment on the date

The lifecycle table says Retired means: "Removed from service. All inference requests return 410 Gone." New deployments are not allowed, and existing deployments do not work. The same page says the API value lifecycleStatus Deprecated means the model is retired and inference returns 410 Gone. The docs stage Deprecated is lifecycleStatus Deprecating, and inference still works. The sentences and the status map are on Azure OpenAI retired models return 410 Gone. These four rows are GA now, so inference still works until June 15, 2027.

The lifecycle page says: "For Global Standard, Data Zone Standard, and Standard deployment types, Microsoft manages automatic upgrades when a model version is retired:" It also says: "Provisioned deployments are NOT auto-upgraded. Provisioned customers must manually migrate to the replacement model." The migration page says: "Standard, Global Standard, and Data Zone Standard (pay-as-you-go) deployments are auto-upgraded on a rolling, region-by-region schedule." And: "Provisioned (PTU) deployments are not auto-upgraded." And: "Batch deployments follow the side-by-side path: deploy the new model, resubmit jobs, retire the old deployment."

Working with models says: "Automatic model updates are only supported for Standard deployment types." OnceCurrentVersionExpired means: "Once the retirement date is reached the model deployment automatically upgrades to the current default version." NoAutoUpgrade means: "The model deployment never automatically upgrades. Once the retirement date is reached the model deployment stops working." And: "null is equivalent to OnceCurrentVersionExpired."

The same page says: "When you select a specific model version for a deployment, this version remains selected until you either choose to manually update yourself, or once you reach the retirement date for the model." Then: "When the retirement date is reached the model will automatically upgrade to the default version at the time of retirement." A deployment you pinned to one of these versions can still move on the retirement date unless versionUpgradeOption is NoAutoUpgrade. NoAutoUpgrade does not move, and the deployment stops. The replacement cell is the target where an automatic upgrade is supported. These four rows have an empty cell, so they do not name that target.

Sources: Azure OpenAI retirement schedule , Foundry Models lifecycle , Working with Azure OpenAI models , Model migration process , checked November 5, 2026.

Find deployments that use these versions

The migration page says: "You also need an inventory of the model deployments your workload uses, including their deployment types (Standard, Provisioned, or Batch)." Working with models says: "To query the current model deployment settings including the deployment upgrade configuration for a given resource use Deployments List." If the value is null, you won't see a versionUpgradeOption property. The model versions page says: "Go to your deployment in the Foundry portal, then open the Details tab to see the current model version." And: "Query the deployments endpoint for your resource. The response includes the model version for each deployment."

The example response puts the model name and version on properties.model. The deployment's own name is the name field next to that object. Match properties.model.name and properties.model.version to the ids in the table. gpt-4o-mini-tts version 2025-12-15 is the dataset id gpt-4o-mini-tts-2025-12-15.

Also search the repository. A hit is a candidate, not proof a deployment still uses that version. How to find every model ID in your codebase covers SDK argument names. The scan page builds a search from dataset ids that have a shutdown date.

Azure PowerShell, list deployments for one account

Get-AzCognitiveServicesAccountDeployment -ResourceGroupName {ResourceGroupName} -AccountName {AccountName}

Deployments List

GET https://management.azure.com/subscriptions/{subscriptionId}/resourceGroups/{resourceGroupName}/providers/Microsoft.CognitiveServices/accounts/{accountName}/deployments?api-version=2023-05-01

ripgrep

rg -n -w -e 'gpt-4o-mini-transcribe-2025-12-15' -e 'gpt-4o-mini-tts-2025-12-15' -e 'gpt-audio-mini-2025-12-15' -e 'gpt-realtime-mini-2025-12-15'

Sources: Working with Azure OpenAI models , Model versions , Model migration process , checked November 5, 2026.

Fine-tuned training and deployment

The schedule and the lifecycle page both say: "Fine-tuned models retire in two phases: training and deployment." And: "Unless explicitly stated, training retires no earlier than the base model retirement date." Both pages say: "For existing customers only. Otherwise, training retirement occurs at base model retirement."

The fine-tune tables on both pages do not list these four versions. gpt-4o-mini-transcribe-2025-12-15: no fine-tune dates are stored, and the fine-tune tables do not list this version. gpt-4o-mini-tts-2025-12-15: no fine-tune dates are stored, and the fine-tune tables do not list this version. gpt-audio-mini-2025-12-15: no fine-tune dates are stored, and the fine-tune tables do not list this version. gpt-realtime-mini-2025-12-15: no fine-tune dates are stored, and the fine-tune tables do not list this version.

A base deployment of these versions uses the retirement date in the first table. Do not borrow a fine-tune date from a different model.

Sources: Azure OpenAI retirement schedule , Foundry Models lifecycle , checked November 5, 2026.

Redeploy before June 15, 2027

Create a deployment of a model you have tested, point traffic at it, and confirm one real audio request before June 15, 2027. Do that for each of the 4 versions in the table.

Working with models describes two ways to change a provisioned deployment once you have chosen a target. In-place migration keeps the deployment name and moves traffic. Multi-deployment migration creates a second deployment, shifts traffic, then deletes the old one. Neither section names a target for these rows.

The batch of these rows is Azure OpenAI shutdowns on June 15, 2027. The Azure OpenAI provider page lists the other stored rows. The April 14, 2027 batch is Azure OpenAI GPT-4.1 and GPT-4o retirement on April 14, 2027. The July 13, 2027 batch is Azure OpenAI gpt-5.1, gpt-5-mini, and gpt-5.2-codex retirement on July 13, 2027.

The pre-migration test checklist is the longer list: evals, rejected fields, token limits, tools, price, and latency. These Azure rows do not store a price.

Sources: Model migration process , Working with Azure OpenAI models , Azure OpenAI retirement schedule , checked November 5, 2026.

Related SunsetDiff pages

All guides

Questions

Which Azure OpenAI versions retire on June 15, 2027?

The dataset lists gpt-4o-mini-transcribe-2025-12-15, gpt-4o-mini-tts-2025-12-15, gpt-audio-mini-2025-12-15, gpt-realtime-mini-2025-12-15. The retirement schedule lists shutdown date 2027-06-15 for each of them. gpt-4o-mini-transcribe-2025-12-15 (GA), gpt-4o-mini-tts-2025-12-15 (GA), gpt-audio-mini-2025-12-15 (GA), gpt-realtime-mini-2025-12-15 (GA).

What is the official replacement?

The recommended replacement is the official replacement stored on each row. gpt-4o-mini-transcribe-2025-12-15: Not published. gpt-4o-mini-tts-2025-12-15: Not published. gpt-audio-mini-2025-12-15: Not published. gpt-realtime-mini-2025-12-15: Not published. The schedule says: "The Replacement column identifies Microsoft's recommended replacement and, where supported, the target for automatic upgrades at retirement." The cell is empty on gpt-4o-mini-transcribe-2025-12-15, gpt-4o-mini-tts-2025-12-15, gpt-audio-mini-2025-12-15, gpt-realtime-mini-2025-12-15.

What does a speech, TTS, chat audio, or realtime app need to change?

gpt-4o-mini-transcribe-2025-12-15 is file transcription at POST /openai/v1/audio/transcriptions. gpt-4o-mini-tts-2025-12-15 is text-to-speech on /audio/speech, and the architecture page says gpt-4o-mini-tts supports prompt-guided style and tone. gpt-audio-mini-2025-12-15 is chat audio. The quickstart says: "Modalities supported in gpt-audio and gpt-audio-mini models include: text, audio, and text + audio." gpt-realtime-mini-2025-12-15 is the Realtime API. The how-to says: "For the Realtime API, use the GA endpoint with /openai/v1 in the URL. Don't use date-based API versions or the api-version query parameter." The replacement cell is empty, so you shortlist the next model and confirm it accepts that same call.

What should I test before June 15, 2027?

Test latency, the voice, the audio format, and the API version on one real call before 2027-06-15. The realtime connection table lists WebRTC at ~100ms, WebSocket at ~200ms, and SIP as Varies. Voices on the audio quickstart: "The following voices are supported for audio out: Alloy, Ash, Ballad, Coral, Echo, Sage, Shimmer, Verse, Marin, and Cedar." Output formats: "The following audio output formats are supported: wav, mp3, flac, opus, pcm16, and aac." Realtime input is PCM 16-bit (pcm16), mono, 24kHz. Transcription input formats: "The audio file object (not file name) to transcribe, in one of these formats: flac, mp3, mp4, mpeg, mpga, m4a, ogg, wav, or webm." Transcription output for gpt-4o-mini-transcribe is json. Chat audio completions use API version 2025-01-01-preview in the gpt-audio-mini sample. Realtime uses /openai/v1 and does not use a date-based api-version.

Why does the schedule also print 2026-12-15 for gpt-realtime-mini 2025-12-15?

Checked on 2026-11-05, the first row says GA and retirement 2027-06-15. A later row says GA and retirement 2026-12-15. Both replacement cells are empty. SunsetDiff keeps the first base-table row. The stored id gpt-realtime-mini-2025-12-15 uses 2027-06-15.

What does Azure return after retirement?

The lifecycle page says a Retired model is removed from service and: "Removed from service. All inference requests return 410 Gone." See Azure OpenAI retired models return 410 Gone.

Will every deployment upgrade itself?

No. The lifecycle page says: "For Global Standard, Data Zone Standard, and Standard deployment types, Microsoft manages automatic upgrades when a model version is retired:" And: "Provisioned deployments are NOT auto-upgraded. Provisioned customers must manually migrate to the replacement model." Working with models says: "Automatic model updates are only supported for Standard deployment types." NoAutoUpgrade means: "The model deployment never automatically upgrades. Once the retirement date is reached the model deployment stops working." null is equivalent to OnceCurrentVersionExpired. These four rows have an empty replacement cell, so they do not name an upgrade target.

How do I find deployments that still use these versions?

Working with models says: "To query the current model deployment settings including the deployment upgrade configuration for a given resource use Deployments List." Read properties.model.name and properties.model.version. The model versions page says: "Go to your deployment in the Foundry portal, then open the Details tab to see the current model version." The migration page says: "You also need an inventory of the model deployments your workload uses, including their deployment types (Standard, Provisioned, or Batch)."

Do fine-tuned deployments of these versions retire on June 15, 2027?

The base retirement date is 2027-06-15. The fine-tune tables do not list these four versions. gpt-4o-mini-transcribe-2025-12-15: no fine-tune dates are stored, and the fine-tune tables do not list this version. gpt-4o-mini-tts-2025-12-15: no fine-tune dates are stored, and the fine-tune tables do not list this version. gpt-audio-mini-2025-12-15: no fine-tune dates are stored, and the fine-tune tables do not list this version. gpt-realtime-mini-2025-12-15: no fine-tune dates are stored, and the fine-tune tables do not list this version. Both pages say: "For existing customers only. Otherwise, training retirement occurs at base model retirement."

Does Azure Government list the same date?

The Azure Government schedule does not list gpt-4o-mini-transcribe, gpt-4o-mini-tts, gpt-audio-mini, or gpt-realtime-mini. The public-cloud schedule lists all four ids on 2027-06-15.

What retires on August 24, 2027?

The schedule lists gpt-audio-1.5 version 2026-02-23, gpt-realtime-1.5 version 2026-02-23, and gpt-5.3-codex version 2026-02-24 as GA with retirement date 2027-08-24 and an empty replacement cell. gpt-audio-1.5-2026-02-23 stores shutdown date 2027-08-24. gpt-realtime-1.5-2026-02-23 stores shutdown date 2027-08-24. gpt-5.3-codex-2026-02-24 stores shutdown date 2027-08-24. See Azure OpenAI shutdowns on August 24, 2027 and Azure OpenAI gpt-5.3-codex, gpt-audio-1.5, and gpt-realtime-1.5 retirement on August 24, 2027. Those ids are not the replacement cell for the 2027-06-15 rows.

More guides

Deprecation alerts

We email you before models shut down. Unsubscribe anytime.

Deprecation alert signup
Providers (optional)