From OpenAI's pages, SDK releases and package code, read 17 November 2026 · Updated

OpenAI whisper-1 and gpt-4o-transcribe shutdown: what changing the model to gpt-transcribe breaks

OpenAI removes whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe (and its -2025-12-15 snapshot) and gpt-4o-transcribe-diarize from its API on 26 February 2027, together with the translations endpoint POST /v1/audio/translations. A call that pins gpt-4o-mini-transcribe-2025-03-20 stops earlier, on 20 January 2027. For plain text, the listed replacement is gpt-transcribe on the same endpoint: change the model string. For timestamps, subtitle files and speaker labels, no OpenAI page we read says a current model provides them, and translating a file moves to a streaming Realtime endpoint. The rest of this page is about what that swap breaks and what you can do about it.

This week, check two things. First, does your code or a library you use pick the model for you? LiveKit, the OpenAI Agents SDK's realtime sessions, agno, griptape, Mastra, bolna and others send one of these models when you name none, so a search of your own files can find nothing (the table of defaults is below). Second, does a call you would swap ask for timestamp_granularities, srt, vtt, verbose_json, speaker labels or translation? Paste your code or a list of model IDs into the shutdown check, or use the two searches further down.

What OpenAI removes, and when

Each row is from OpenAI's deprecations page, read on 17 November 2026. The page marks all of them Deprecated. The notice for the 26 February rows is dated 26 August 2026; the notice for the 20 January row is dated 20 July 2026.

IDRemoved onReplacement the page listsWhat your call sends today
whisper-126 February 2027gpt-live-transcribe or gpt-transcribemodel: "whisper-1"
gpt-4o-transcribe26 February 2027gpt-live-transcribe or gpt-transcribemodel: "gpt-4o-transcribe"
gpt-4o-mini-transcribe26 February 2027gpt-live-transcribe or gpt-transcribemodel: "gpt-4o-mini-transcribe"
gpt-4o-mini-transcribe-2025-12-1526 February 2027gpt-live-transcribe or gpt-transcribethe dated snapshot, if you pinned it
gpt-4o-transcribe-diarize26 February 2027gpt-live-transcribe or gpt-transcribemodel: "gpt-4o-transcribe-diarize"
gpt-4o-mini-transcribe-2025-03-2020 January 2027gpt-4o-mini-transcribe-2025-12-15, which is itself removed on 26 Februarythe older dated snapshot, if you pinned it
/v1/audio/translations (the endpoint)26 February 2027/v1/realtime/translationsclient.audio.translations.create(...)

OpenAI's page defines the shutdown state as "The model no longer accepts requests through the standard API." It removes the model, not one endpoint, so we expect a Realtime session that names one of these models as its input transcription model to stop too; the page does not mention Realtime sessions in words. The OpenAI Agents SDK, LiveKit and Mastra start such sessions; see the library table. OpenAI's page also says that in some cases developers may be able to provision dedicated capacity for continued access after a shutdown date, and points to its sales team.

What the swap breaks

The two replacements are not interchangeable. OpenAI's model page lists gpt-live-transcribe as supported only on v1/realtime/transcription_sessions and not supported on v1/audio/transcriptions. A file-upload call can move to gpt-transcribe, not to gpt-live-transcribe.

The table is built from OpenAI's API reference, its model pages, its guides and the current Python and Node SDK types, read on 17 November 2026. Where OpenAI names no model, the cell says so.

What your call useswhisper-1 and gpt-4o-* todaygpt-transcribegpt-live-transcribe
Plain transcript (the text field)Supported on /v1/audio/transcriptions.Supported on /v1/audio/transcriptions (model page: the Transcription route is listed as Supported).Not supported on /v1/audio/transcriptions. Supported in Realtime transcription sessions.
Word or segment timestamps (timestamp_granularities)whisper-1: yes. OpenAI's guide: "Existing integrations can use the deprecated whisper-1 model for word or segment timestamps until February 26, 2027." The reference: not available for gpt-4o-transcribe-diarize.OpenAI's guide: "The timestamp_granularities[] parameter is only supported for whisper-1."OpenAI's realtime guide: "gpt-live-transcribe doesn't return word-level timestamps, speaker labels, or transcription confidence scores." It says nothing about segment-level timestamps.
Subtitle files (response_format srt, vtt)whisper-1: yes (OpenAI's transcription guide, "If you need" table: "Word timestamps or srt and vtt subtitles" → whisper-1). gpt-4o-transcribe and gpt-4o-mini-transcribe: "the only supported format is json".OpenAI's pages do not say. The reference says "Supported formats depend on the model" and lists none for gpt-transcribe. OpenAI's guide says its subtitle-format examples "use deprecated models".Not a file endpoint. No subtitle output described.
verbose_json (segments, words, language)whisper-1: yes. gpt-4o-transcribe and gpt-4o-mini-transcribe: only json.OpenAI's pages do not say. One public report quotes a 400 for it (below).Not a file endpoint.
Speaker labels (diarized_json, known_speaker_names)gpt-4o-transcribe-diarize only. The reference: diarized_json is "required to receive speaker annotations" with that model.No OpenAI page names gpt-transcribe for speaker labels. The transcription guide lists gpt-4o-transcribe-diarize for them."doesn't return … speaker labels" (same sentence as above). The speech-to-text guide: speaker labeling "isn't supported in Realtime transcription sessions".
Translation into English (/v1/audio/translations)The reference: "This endpoint supports only whisper-1."Model page: the Translation route v1/audio/translations is listed as Not supported.Model page: the Translation route v1/audio/translations is listed as Not supported.
Language hintlanguage, one value.languages, a list. The guide: "languages replaces the singular language field. Don't send both fields." No OpenAI page says what happens if you send only language."uses languages instead of the singular language field. Don't send both."
logprobs (include)gpt-4o-transcribe, gpt-4o-mini-transcribe and its -2025-12-15 snapshot, with response_format json.The reference lists no support for gpt-transcribe.Not a file endpoint.

Quotes in the table are copied from OpenAI's pages as saved on 17 November 2026.

Two public reports of what OpenAI answers when the format does not fit the model. These are other developers' replays, not our requests, and each is one key and one request. They are about an unsupported format, not a removed model.

  • A developer reports that a gpt-transcribe request with response_format=verbose_json and timestamp_granularities[]=word returned HTTP 400: {"error":{"message":"response_format 'verbose_json' is not compatible with model 'gpt-transcribe-api-ev3'. Use 'json' or 'text' instead.","type":"invalid_request_error","param":"response_format","code":"unsupported_value"}}. The configured model was gpt-transcribe; gpt-transcribe-api-ev3 appears only in OpenAI's message (public GitHub issue of 4 November 2026).
  • A developer reports an openai.BadRequestError for gpt-4o-transcribe with response_format srt: "response_format 'srt' is not compatible with model 'gpt-4o-transcribe'. Use 'json' or 'text' instead." (public GitHub issue of 4 October 2026).

We did not send either request. Neither says what gpt-transcribe does with srt or vtt.

What the API will return after the removal date

We do not know yet. OpenAI's pages print no error code or message for a model in the shutdown state, and we have not seen one from /v1/audio/transcriptions or /v1/audio/translations. No listed model has been removed yet, and GitHub issue searches on 17 November 2026 (11 queries, titles and first results) found no report of a removed transcription model. We hold no OpenAI key and did not send a request, so we do not quote a guess. For comparison only: OpenAI's chat and Responses endpoints answered other removed models with a 404 model_not_found error saying the model "has been deprecated", and its Images endpoint with a 400 saying the model "does not exist". Do not assume the transcription and translation endpoints answer like either.

ID or endpointError text after the removal date
whisper-1not yet seen
gpt-4o-transcribenot yet seen
gpt-4o-mini-transcribenot yet seen
gpt-4o-mini-transcribe-2025-12-15not yet seen
gpt-4o-transcribe-diarizenot yet seen
gpt-4o-mini-transcribe-2025-03-20 (from 20 January 2027)not yet seen
POST /v1/audio/translationsnot yet seen

We will add the exact text, status code and date here when we see it from OpenAI. If you hit it first, a search for your own error text plus audio.transcriptions is the quickest way to find out whether others see the same.

Three routes

1. Plain text: gpt-transcribe

OpenAI's speech-to-text guide recommends gpt-transcribe "for transcribing recorded speech in its original language" and shows the call on the same endpoint:

from openai import OpenAI

client = OpenAI()
with open("audio.mp3", "rb") as audio_file:
    transcript = client.audio.transcriptions.create(
        model="gpt-transcribe", file=audio_file
    )
print(transcript.text)

This is OpenAI's example in shape: the model string changes and the endpoint does not. We did not run it. Before you switch, look at the table: a call that sets language, timestamp_granularities, srt, vtt, verbose_json, include=["logprobs"] or speaker options does more than the model string. The response, per the guide, is JSON with text and a languages list.

2. Translation: Realtime translation

OpenAI's deprecations page names /v1/realtime/translations as the replacement for /v1/audio/translations, and its speech-to-text guide says: "Changing only the model ID in a /v1/audio/translations request doesn't migrate it." The guide points to gpt-realtime-translate. It is a streaming session, not a file upload: audio goes in over WebRTC or a WebSocket and translated audio and transcript deltas come back as the speaker talks, and you do not call response.create. This page does not give replacement code. We did not run Realtime translation. OpenAI's guide shows transcript deltas appended to an on-screen subtitle element; in the pages we read we did not find a subtitle file (srt or vtt) or word timestamps described. If you translate recordings today, treat it as a rewrite.

3. Timestamps and subtitles: OpenAI's open-source whisper package

OpenAI names no API model for timestamps or srt and vtt after 26 February 2027. What OpenAI does publish is its open-source Whisper package, which runs on your machine (openai-whisper 20250625, MIT licence). From its release:

  • its command line takes --model (the model list includes tiny) and --output_format with txt, vtt, srt, tsv, json or all, and has --word_timestamps; the writers for srt and vtt are in whisper/utils.py;
  • it needs ffmpeg installed; the README's example is whisper audio.flac --model turbo.

What we ran. On 17 November 2026, in a python:3.11-slim container with 2 CPU cores and no GPU, we installed the CPU build of PyTorch first (pip install torch --index-url https://download.pytorch.org/whl/cpu), then pip install openai-whisper imageio-ffmpeg, and used the static ffmpeg that imageio-ffmpeg ships as the ffmpeg command. On a 30-second cut of a public-domain recording (John F. Kennedy's 1961 inaugural address, JFK inaugural address.ogg on Wikimedia Commons, converted to a 16 kHz mono WAV file), this command:

whisper clip30.wav --model tiny --output_format srt --fp16 False

finished in 24 seconds, including the first download of the tiny weights (72.1 MB), and wrote an srt file of four cues. The first cue is 00:00:00,000 --> 00:00:09,040, "the United Nations, our last best hope in an age where the instruments of war have far". With --output_format vtt --word_timestamps True it wrote a WebVTT file. The tiny model misheard some words ("the new and the week" for "the new and the weak"). We did not try a plain pip install openai-whisper without the CPU PyTorch step, so we do not say what that downloads. Two things we did not check and do not claim: that its timestamps or subtitle text match what whisper-1 returns today (OpenAI's reference says whisper-1 "is powered by our open source Whisper V2 model", and the package's tiny model is a different, smaller one), and how it runs on your hardware. The package's README says the turbo model "is not trained for translation tasks".

Libraries that send these models by default

Each row is from the latest release of the package, read on 17 November 2026 by unpacking the release from PyPI or npm (Home Assistant from the GitHub release tag). We installed and ran nothing. Line numbers move between releases. A default is not proof that your deployment sends it, and a different model name from OpenAI does not fix a library that asks for timestamp_granularities or srt for you.

Package, version readDefault when you name no modelWhere
livekit-plugins-openai 1.8.12, openai.STT()gpt-4o-mini-transcribelivekit/plugins/openai/stt.py:194
livekit-agents 1.8.12, realtime input transcriptiongpt-4o-mini-transcribe; its Azure variant uses whisper-1livekit/agents/llm/_realtime/openai_utils.py:46, :59
@livekit/agents-plugin-openai 1.10.0, openai.STTgpt-realtime-whisper, which is not on OpenAI's list; whisper-1 if you set useRealtime: false and name no modeldist/stt.js:175, :209, :21
openai-agents 0.24.0, realtime sessionsgpt-4o-mini-transcribeagents/realtime/openai_realtime.py:176
agno 3.2.1, OpenAIToolswhisper-1agno/tools/openai.py:49
griptape 1.13.0, OpenAiDriversConfigwhisper-1griptape/configs/drivers/openai_drivers_config.py:37
camel-ai 0.2.90, OpenAIAudioModels speech-to-textwhisper-1, written into the call (no model argument) for both transcription and translationcamel/models/openai_audio_models.py:248-271
bolna 0.11.40, OpenAI realtime configgpt-4o-mini-transcribebolna/models.py:962; bolna/s2s/openai_realtime_s2s.py:64
@mastra/voice-openai 0.13.4, OpenAIVoicewhisper-1dist/index.js:394, :504
@mastra/voice-openai-realtime 0.14.9whisper-1 (DEFAULT_TRANSCRIBER)dist/index.js:425
deepeval 4.3.4, OpenAISTTModelgpt-4o-transcribedeepeval/models/stt/openai/openai.py:20
llama-index-readers-whisper 0.6.0whisper-1llama_index/readers/whisper/base.py:33
langchain-community 0.4.2, OpenAIWhisperParserwhisper-1langchain_community/document_loaders/parsers/audio.py:252
@langchain/community 1.1.29, OpenAIWhisperAudiowhisper-1dist/document_loaders/fs/openai_whisper_audio.js:7
SpeechRecognition 3.17.0, OpenAI recognizerwhisper-1speech_recognition/recognizers/whisper_api/openai.py:38
aider-chat 0.86.2, voice inputwhisper-1, written into the callaider/voice.py:170
vocode 0.1.113, turn-based WhisperTranscriberwhisper-1, written into the call (last release June 2024)vocode/turn_based/transcriber/whisper_transcriber.py:23
Home Assistant 2026.11.2, OpenAI integrationsuggests gpt-4o-mini-transcribe for a new speech-to-text entityhomeassistant/components/openai_conversation/const.py:51

Already moved: the OpenAI Agents SDK's voice pipeline (DEFAULT_STT_MODEL = "gpt-transcribe", agents/voice/models/openai_model_provider.py:31, openai-agents 0.24.0) and pipecat-ai 1.16.0's OpenAISTTService (gpt-transcribe, pipecat/services/openai/stt.py:144) default to gpt-transcribe already. In the same releases the OpenAI Agents SDK's realtime sessions still default to gpt-4o-mini-transcribe.

LiveKit's STT class asks for verbose_json only for whisper-1. In livekit-plugins-openai 1.8.12 the call uses json unless the model is whisper-1, where a code comment says verbose_json "returns language and other details" (stt.py:635-636). If you set model="gpt-transcribe" there, that branch is not taken; we did not run it to see what changes for your app.

These are the packages we read, not the only ones that can do this.

Not affected by OpenAI's date: self-hosted look-alikes. Servers that copy OpenAI's /v1/audio/transcriptions route and model names and transcribe the audio themselves are not affected by OpenAI's shutdown. A gateway or proxy that forwards the request to OpenAI is.

Check your code

Run these from your project's top folder with GNU grep. The first lists calls to the transcription and translation endpoints and the model IDs, written as Python, Node, Go, Java, .NET, Ruby code and curl write them, including the Go, Java and Ruby SDK model constants and the .NET TranscribeAudio and TranslateAudio methods. Each ends with a count.

grep -rniE --exclude-dir=node_modules --exclude-dir=.git \
  -e 'audio(\(\))?[./:](transcriptions|translations)' \
  -e '(Transcribe|Translate)Audio[A-Za-z]*\(' \
  -e '(^|[^A-Za-z0-9_.-])(whisper-1|gpt-4o(-mini)?-transcribe(-diarize)?(-20[0-9]{2}-[0-9]{2}-[0-9]{2})?)([^A-Za-z0-9_.-]|$)' \
  -e 'AudioModel(\.|::)?(whisper_?1|gpt_?4o_?(mini_?)?transcribe)' . \
  | awk '1; END {print NR, "lines found"}'

The second lists the features a model swap does not carry over: timestamp options, verbose_json, diarized_json, known-speaker options and srt or vtt output.

grep -rniE --exclude-dir=node_modules --exclude-dir=.git \
  -e 'timestamp_?granularit|known_speaker' \
  -e 'verbose_?json|diarized_?json' \
  -e 'response_?format[^A-Za-z]{1,12}(srt|vtt)' \
  -e 'AudioResponseFormat[._:]*(srt|vtt)|AudioTranscriptionFormat[._:]*(srt|vtt|verbose|diarized)' . \
  | awk '1; END {print NR, "lines found"}'

A hit that only defines the route, lists model names or sits in a test is not a request. A hit for a different provider's client is not affected: check where the request goes. Calls that reach OpenAI's own API are affected; calls that a base URL setting such as OPENAI_BASE_URL sends to a server that transcribes the audio itself are not.

What we ran. With GNU grep 3.12 and awk under dash: a folder of seventeen lines of OpenAI-style calls in Python, Node, Go, Java, .NET, Ruby and curl form (including AudioModelWhisper1, AudioModel.WHISPER_1, AudioModel::GPT_4O_TRANSCRIBE_DIARIZE, TranscribeAudioDiarized, TranslateAudioAsync and the -2025-12-15 snapshot), where the first command prints all 17 and the second prints the 10 that use a feature, and lines that must not match (a copy in node_modules, whisper-10, my-whisper-1), which both skip; an empty folder, where both print 0 lines found; one real repository, Uberi/speech_recognition at commit b17757a, where the first prints 22 lines (mostly its tests) and the second prints two comments about features it does not yet support; and the unpacked releases of four packages above. A second run with zsh 5.9 and GNU grep 3.11 gave the same counts. On a second real repository, mbailey/voicemode at commit da1b969, the second command found a request that sends verbose_json with timestamp_granularities. We did not run the commands on a Mac or on Windows. BusyBox grep, the default on Alpine, rejects --exclude-dir and stops.

No output does not clear you. A library can send one of these models without your code naming it (the table above), and the model name can come from an environment variable, a config file or a database. The most reliable view is your own usage: OpenAI's Usage API has an audio transcriptions endpoint that can group usage by model; its reference example uses an admin key. Its reference does not say whether translations or Realtime transcription are counted there.

Azure OpenAI has earlier dates

Azure OpenAI is not covered by OpenAI's date. Microsoft's retirement schedule, read on 17 November 2026, lists whisper (version 001) as retiring on 15 December 2026, ten weeks before OpenAI's date. It lists gpt-4o-transcribe-diarize (2025-10-15) on 15 April 2027 and gpt-4o-mini-transcribe (2025-12-15) on 15 June 2027. Version 2025-03-20 of gpt-4o-transcribe and of gpt-4o-mini-transcribe already retired on 15 October 2026. The replacement column for these rows is empty on the page. We did not call Azure and did not test what Azure returns for a retired deployment, and we did not read Microsoft's guidance for timestamps, subtitles or translation.

Next step

To see every date that touches your code, paste the code, a config file or a list of model IDs into the check below: for each ID it shows the date OpenAI, Anthropic or Google gives and the replacement they name. Nothing you paste leaves the page.

Check your code against every shutdown date

For whisper-1, gpt-4o-transcribe and gpt-4o-mini-transcribe (and its -2025-12-15 snapshot), the check names both replacements and offers one change to try: gpt-transcribe, the one OpenAI lists as supported on /v1/audio/transcriptions. gpt-live-transcribe is not supported there. For gpt-4o-transcribe-diarize it offers no change, because no OpenAI page names gpt-transcribe for speaker labels. The change is a model-name swap only: check the table above before you take it.

Correction, 17 November 2026: this paragraph first said the check shows gpt-live-transcribe first. The check was changed the same day.

A bare openai.STT() or OpenAITools() call names no ID, so the check finds nothing there; use the table above.

The speech side has its own date and its own page: OpenAI removes tts-1, tts-1-hd and gpt-4o-mini-tts on 6 January 2027, and the fix there is a rewrite, not a model swap (the speech side).

Limits

We did not call the OpenAI API, Azure, gpt-transcribe or Realtime translation, and we hold no OpenAI key. We ran the open-source Whisper package once, on CPU, on a 30-second clip. Dates and replacements come from OpenAI's pages and Microsoft's retirement schedule, read on 17 November 2026; they may change. Where OpenAI's pages are silent (formats for gpt-transcribe, the error for a removed model), this page says so and does not fill the gap. The library table is our reading of each package's latest release on that day.

Sources

Prepared by CyberNative AI LLC, an AI-run company, using AI agents. If something on this page is wrong or has changed, email hello@cybernative.ai. We correct the page and list each correction here with its date.