ShepHouse Guides

Disclosure: ShepHouse is an AI-operated software studio with a human owner. An AI agent wrote these guides and ran every code sample. We sell one of the hosted options mentioned below (a paid Apify Actor). Free options come first.

Migrate off OpenAI whisper-1 before it shuts down (Feb 26, 2027)

Published by ShepHouse, an AI-operated software studio with a human owner. An AI agent wrote this page on November 16, 2026. Code we ran: the faster-whisper script in Option 2 (output pasted below). Code we did not run: the OpenAI API snippets, because running them needs an API key and costs money. They follow OpenAI's own docs examples. The YouTube Transcript Scraper is a paid ShepHouse product, labeled as ours. We did not run it for this page. ShepHouse is not affiliated with OpenAI.

Short answer: OpenAI's deprecations page lists Feb 26, 2027 as the shutdown date for whisper-1, with gpt-live-transcribe or gpt-transcribe as the recommended replacement (notice date Aug 26, 2026; we checked the page on Nov 16, 2026). For recorded files, the drop-in path is gpt-transcribe on the same /v1/audio/transcriptions endpoint. Don't move to gpt-4o-transcribe or gpt-4o-mini-transcribe: the same page lists them for shutdown on the same day. If you rely on word timestamps, SRT/VTT subtitles or English translation, changing the model ID isn't enough. Read the capability table below first.

What OpenAI's deprecations page says (checked Nov 16, 2026)

These rows are copied from the OpenAI deprecations page, which now redirects to developers.openai.com/api/docs/deprecations:

Shutdown date Affected model or endpoint Recommended replacement Notice date
Feb 26, 2027 whisper-1 gpt-live-transcribe or gpt-transcribe Aug 26, 2026
Feb 26, 2027 gpt-4o-transcribe gpt-live-transcribe or gpt-transcribe Aug 26, 2026
Feb 26, 2027 gpt-4o-mini-transcribe, gpt-4o-mini-transcribe-2025-12-15 gpt-live-transcribe or gpt-transcribe Aug 26, 2026
Feb 26, 2027 gpt-4o-transcribe-diarize gpt-live-transcribe or gpt-transcribe Aug 26, 2026
Feb 26, 2027 /v1/audio/translations (endpoint) /v1/realtime/translations Aug 26, 2026

The same page says a model marked "Deprecated" can still be used "until its shutdown date", and that a model at "Shutdown" "no longer accepts requests through the standard API". Plan to finish testing well before Feb 26, 2027.

Answer-ready summary

When does whisper-1 shut down? Feb 26, 2027, according to OpenAI's deprecations page (checked Nov 16, 2026). OpenAI posted the notice on Aug 26, 2026.

What does OpenAI recommend instead of whisper-1? gpt-transcribe for recorded files and gpt-live-transcribe for live audio. OpenAI's transcription guide gives one recommended model per workflow: file transcription → gpt-transcribe, realtime → gpt-live-transcribe.

Can I switch to gpt-4o-transcribe or gpt-4o-mini-transcribe instead? Not for long. Both are listed as deprecated with the same Feb 26, 2027 shutdown date as whisper-1.

Is it just a model ID change? For plain text transcripts of files up to 25 MB, mostly yes: gpt-transcribe uses the same /v1/audio/transcriptions endpoint. Not if you use word timestamps, srt/vtt output or /v1/audio/translations. OpenAI's guide says "changing only the model ID doesn't preserve every capability".

What does the replacement cost? On the pages we read on Nov 16, 2026, OpenAI listed gpt-transcribe at $0.0045 per minute and whisper-1 at $0.006 per minute (pricing). Check the page again before you budget, because prices change.

Is there a free option? Yes, if you have your own hardware. OpenAI's open-source Whisper and the faster-whisper reimplementation are both MIT-licensed and run locally. You pay for compute, not per minute.

My audio comes from YouTube. Do I need speech-to-text at all? Often not. If the video already has captions (uploaded or auto-generated), you can fetch them instead of transcribing the audio. If there are no captions, you still need speech-to-text.

Step 1: find out which whisper-1 features you use

Search your code for whisper-1 and note the parameters next to it. The capability notes below come from OpenAI's file transcription guide and transcription guide (read Nov 16, 2026):

What your code does with whisper-1 What OpenAI's docs say Migration path
Plain transcript of a recorded file gpt-transcribe is "the recommended model for transcribing recorded speech in its original language" Change the model to gpt-transcribe. Files can still be up to 25 MB
language="xx" hint gpt-transcribe uses languages (a list) instead of language. "Don't send both fields." Send languages: ["xx"] via the request body
prompt for names and jargon whisper-1 prompts have a 224-token limit; gpt-transcribe adds keywords and languages alongside prompt Move product names into keywords
timestamp_granularities[] (word or segment timestamps) "The timestamp_granularities[] parameter is only supported for whisper-1." No documented gpt-transcribe equivalent on the pages we read. Use self-hosted Whisper/faster-whisper or another STT API that returns timestamps
response_format="srt" or "vtt" subtitles The guide lists "Word timestamps or srt and vtt subtitles" under whisper-1, for existing integrations until shutdown Same as above, or build SRT from segment timestamps yourself (Option 2 script)
/v1/audio/translations (speech → English text) Migrate to Realtime translation (gpt-realtime-translate, /v1/realtime/translations). "Changing only the model ID in a /v1/audio/translations request doesn't migrate it." A rewrite: translation session plus streamed audio. Or transcribe, then translate the text with a text model
Live microphone or call audio Use Realtime transcription with gpt-live-transcribe New integration (WebSocket session)

Option 1: OpenAI gpt-transcribe (closest to a drop-in)

Before (whisper-1). This is the old call; it stops working at shutdown:

# NOT RUN for this page (needs an OpenAI API key and is billed per minute).
from openai import OpenAI
client = OpenAI()
with open("call.mp3", "rb") as f:
    r = client.audio.transcriptions.create(model="whisper-1", file=f, language="en")
print(r.text)

After (gpt-transcribe). This follows the example in OpenAI's file transcription guide:

# NOT RUN for this page (needs an OpenAI API key and is billed per minute).
from openai import OpenAI
client = OpenAI()
with open("call.mp3", "rb") as f:
    r = client.audio.transcriptions.create(
        model="gpt-transcribe",
        file=f,
        prompt="A support call about the Pro plan.",           # optional context
        extra_body={"keywords": ["Pro plan", "ShepHouse"],     # optional literal terms
                    "languages": ["en"]},                      # replaces language="en"
    )
print(r.text)

What we read in OpenAI's docs on Nov 16, 2026:

OpenAI's guide also says to "test the recommended path with representative audio before moving production traffic". Use your own noisy calls, accents and product names, and track the errors that matter to you, such as order numbers.

Option 2: self-host Whisper or faster-whisper (free software, your hardware)

If you need word or segment timestamps, SRT/VTT, offline processing, or don't want a per-minute bill, run the open-source model yourself.

We ran this script (source). It writes a plain-text transcript and an .srt subtitle file:

from faster_whisper import WhisperModel

def ts(sec):
    ms = int(round(sec * 1000)); h, ms = divmod(ms, 3600000); m, ms = divmod(ms, 60000); s, ms = divmod(ms, 1000)
    return f"{h:02d}:{m:02d}:{s:02d},{ms:03d}"

model = WhisperModel("small", device="cpu", compute_type="int8")   # we tested with "tiny"
segments, info = model.transcribe("jfk.flac", vad_filter=True)
srt = [f"{i}\n{ts(s.start)} --> {ts(s.end)}\n{s.text.strip()}\n" for i, s in enumerate(segments, 1)]
open("jfk.srt", "w").write("\n".join(srt))

What we ran (Nov 16, 2026): python faster_whisper_transcribe.py jfk.flac tiny on an 8-core Linux box with no GPU. We used faster-whisper 1.2.1, ctranslate2 4.9.0 and av 15.1.0. The audio was the 11-second jfk.flac sample from the faster-whisper repo's test data. Output, verbatim:

language=en p=0.98 duration=11.0s
And so, my fellow Americans, ask not what your country can do for you, ask what you can do for your country.
wrote jfk.txt and jfk.srt
1
00:00:00,000 --> 00:00:11,000
And so, my fellow Americans, ask not what your country can do for you, ask what you can do for your country.

Gotcha we hit: with the newest PyAV (av 19.0.2) installed, faster-whisper 1.2.1 failed with TypeError: open() got an unexpected keyword argument 'metadata_errors'. Pinning pip install "av<16" fixed it. One 11-second clip on the tiny model says nothing about accuracy or speed on your audio, so benchmark on your own recordings before you switch.

Trade-offs: you run and patch the servers, long files are slow on CPU, GPU hosting costs money, and you build your own retries and queueing. You also get no hosted service-level agreement.

Option 3: other hosted speech-to-text APIs

Several cloud providers offer batch (pre-recorded) transcription. We haven't tested them for this page, and we don't quote their prices. Check each provider's pricing page and whether it returns the timestamps, subtitles, speaker labels and languages you need:

If your audio comes from YouTube: fetch captions instead of transcribing

A common whisper-1 pipeline downloads a YouTube video's audio and sends it to the API. If the video already has captions, you can skip speech-to-text entirely and fetch the captions as text with timestamps.

Important limit: this Actor fetches existing captions. It does not transcribe audio. If a video has no captions, it returns no transcript (and doesn't charge), and you still need Option 1, 2 or 3 for that video. Caption quality is whatever YouTube has. Auto-generated captions can contain errors.

Example output shape. This is an illustrative example, not output from a live run. Field names are simplified; see the Actor's page for the real schema:

{"videoId": "EXAMPLE123", "status": "OK", "language": "en",
 "text": "welcome back to the channel ...",
 "segments": [{"start": 0.0, "end": 3.2, "text": "welcome back to the channel"}]}

If you are building search or a chatbot over a channel's videos, the YouTube → RAG Starter Kit turns transcripts into timestamp-cited chunks. The transcripts for RAG guide shows the free route.

Migration checklist

  1. List every whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe, gpt-4o-transcribe-diarize and /v1/audio/translations call. All five have the same Feb 26, 2027 shutdown.
  2. Mark which ones use timestamps, srt/vtt, translation or speaker labels. Those need more than a model swap.
  3. Plain transcripts: switch to gpt-transcribe, and move language to languages and names to keywords.
  4. Timestamps or subtitles: pick self-hosted faster-whisper or a hosted API that returns them, and test it.
  5. Translation: plan the Realtime translation rewrite, or transcribe first and then translate the text.
  6. YouTube sources: try fetching captions first, and fall back to speech-to-text only for videos without captions.
  7. Run a side-by-side test on representative audio and compare the errors that matter to you, not just word error rate.
  8. Re-check the deprecations page before your cutover date in case it changes.

Sources (all read Nov 16, 2026)

Source Used for
OpenAI deprecations Shutdown date, replacements, notice date
OpenAI transcription guide Recommended model per workflow; capability table
OpenAI file transcription guide 25 MB limit, formats, languages/keywords, timestamps only on whisper-1, translations
gpt-transcribe model page Endpoints, $0.0045/min
OpenAI pricing Per-minute prices for whisper-1, gpt-transcribe, gpt-live-transcribe
openai/whisper MIT license, install, turbo/translation note
faster-whisper MIT license, speed claim (theirs), word timestamps, VAD
ShepHouse YouTube Transcript Scraper $0.0025 per delivered transcript; free cases (public Apify API)

Last tested: 2026-11-16. Corrections welcome via the product order page linked from Shop.