whisper-1 deprecation: replacements and costs before the 26 Feb 2027 cut-off

Summary. OpenAI will remove whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe, gpt-4o-transcribe-diarize and the /v1/audio/translations endpoint on 26 February 2027, a date still shown on its deprecations page on 1 November 2026. The named replacements are gpt-transcribe at $0.0045 a minute for recorded files and gpt-live-transcribe at $0.017 a minute for live audio; translation moves to gpt-realtime-translate on /v1/realtime/translations at $0.034 a minute. Moving whisper-1 or gpt-4o-transcribe files saves 25%. Almost everything else gets dearer or loses a feature: gpt-4o-mini-transcribe files cost 50% more, live captioning can rise almost sixfold, and translating a recorded file becomes a streaming session at 5.7 times the price. OpenAI's guides document word timestamps, subtitles and speaker labels only on the models being removed. For Indian languages, OpenAI publishes no per-language accuracy. In our 100,000-minute example with speaker labels, AssemblyAI Universal-2 ($283), Azure Speech batch ($300), Google Chirp 3 dynamic batch ($300) and Deepgram Nova-3 ($520) undercut today's $600, and Sarvam Saaras V4, with 22 Indian languages, comes to about $621. Inventory every call site now, test on real Indian audio in December, and cut over before mid-February.

About this article. This guide is for teams running speech-to-text on OpenAI. Dates, prices and capabilities were checked on 1 November 2026 against OpenAI's deprecations page, guides, model pages, API reference and pricing, plus Deepgram, AssemblyAI, Google Cloud, Microsoft and Sarvam pages. Vendors change prices often, so re-check before you commit; this is not legal advice, and no vendor paid for inclusion. Researched with AI assistance and reviewed by our editorial team.

What OpenAI removes on 26 February 2027

OpenAI notified affected developers on 26 August 2026 and lists the shutdown in the transcription block of its deprecations page. Unlike the January text-to-speech cutoff, this notice includes an endpoint, not only model names. Prices below come from OpenAI's pricing and model pages.

One snapshot leaves earlier. The legacy audio notice of 20 July 2026 removes gpt-4o-mini-transcribe-2025-03-20 on 20 January 2027 and points to the 2025-12-15 snapshot, which then goes on 26 February. Pinning that snapshot buys five weeks, not a fix.

OpenAI's minimum notice for generally available models is six months, which this notice meets exactly. Some developers may be able to arrange dedicated capacity through sales; treat that as a negotiation for large accounts, not a plan. To track dates in code, call GET /v1/models and read each model's optional shutdown_date field.

Voice teams face three OpenAI audio deadlines in under eight weeks: Speech API models on 6 January, legacy Realtime and audio models on 20 January, and transcription on 26 February. Our OpenAI TTS and Realtime shutdown guide covers the first two.

Azure OpenAI runs its own calendar. Microsoft's schedule retires whisper 001 on 15 December 2026, well before OpenAI's date, while gpt-4o-transcribe-diarize 2025-10-15 runs to 15 April 2027 and gpt-4o-mini-transcribe 2025-12-15 to 15 June 2027.

Three migrations, not one

The notice reads like a rename. In practice it means three different code changes.

Recorded files: whisper-1 to gpt-transcribe

gpt-transcribe, released on 28 July 2026, runs on the same /v1/audio/transcriptions endpoint, so a basic upload changes only the model field. The 25 MB file limit stays. Our gpt-transcribe and gpt-live-transcribe comparison covers accuracy and cost. The differences sit in parameters and outputs:

• Language hints are plural. gpt-transcribe takes languages, a list, instead of language, and OpenAI says not to send both. For Hinglish calls, send both Hindi and English. The API rejects unsupported or badly formatted codes, so test each code you rely on.

• Context is richer. prompt takes free-form context and a new keywords list takes literal terms such as product names. whisper-1 prompts are capped at 224 tokens.

• Detected languages come back. The response includes a languages array, empty when the model cannot decide.

• Streaming works on files. Set stream=true to receive transcript.text.delta events while a recorded file is processed. whisper-1 ignores streaming.

• Confidence data goes away. Log probabilities are returned only by gpt-4o-transcribe and the mini model. If they route low-confidence segments to human review, plan a replacement signal.

Live audio: gpt-4o-transcribe sessions to gpt-live-transcribe

gpt-live-transcribe runs only on Realtime transcription sessions (/v1/realtime/transcription_sessions). You connect to wss://api.openai.com/v1/realtime?intent=transcription, send session.update with the session type set to transcription, and stream audio; OpenAI's example uses 24 kHz PCM. The model does not support server_vad or semantic_vad, so turn_detection must be omitted or null and your client decides when to send input_audio_buffer.commit. And a new delay setting runs from minimal to xhigh; OpenAI says higher delay can improve word error rate but advises benchmarking instead of assuming a fixed timing per level.

If latency matters less than price, gpt-transcribe can also transcribe committed turns over a Realtime WebSocket. OpenAI lists one price for it, $0.0045 a minute, and it returns detected languages where the live model does not. The guide also describes gpt-transcribe handling input transcription inside a Realtime voice-agent session, so voice agents that set gpt-4o-transcribe as the input transcription model need that field changed.

A third model, gpt-realtime-whisper at $0.017 a minute, is not on the deprecation list, but OpenAI does not recommend it as the starting model for a new integration.

Translation: /v1/audio/translations to gpt-realtime-translate

This is the largest change. The old endpoint took a finished recording and returned English text. OpenAI's guide is explicit that changing the model ID alone does not migrate a request. The replacement is a translation session that streams source audio in and returns translated audio plus transcript deltas. It needs WebRTC or a WebSocket with base64 24 kHz PCM16, continuous audio including silence, one session per target language, and a session.close call followed by waiting for session.closed.

For a batch job that translated stored files into English text, a live interpreter model is the wrong shape. Cheaper patterns are transcribing with gpt-transcribe and translating the text, or using a service with a translate mode.

The feature gaps: timestamps, subtitles, speakers and confidence

The deprecation table lists replacements, not feature parity. OpenAI's file transcription guide says its examples for speaker labels, word timestamps and subtitle formats use deprecated models, and that timestamp_granularities[] works only with whisper-1. The Realtime guide adds that gpt-live-transcribe returns no word-level timestamps, speaker labels or confidence scores. The vendor column below follows each vendor's pricing and documentation pages.

OpenAI could add these features before February. As of 1 November 2026 it had not documented them, so subtitle pipelines, meeting-notes apps and QA tools should treat speaker labels and timestamps as a separate procurement decision.

There is one low-change exit. OpenAI's API reference says whisper-1 is powered by its open-source Whisper V2 model, and the open-source Whisper repository is MIT-licensed. Self-hosting that model keeps timestamps and English translation, at the cost of running GPUs yourself. Deepgram also hosts Whisper Large for pre-recorded audio at $0.0048 a minute; check which outputs it returns.

Accuracy for Indian languages

OpenAI. The gpt-transcribe and gpt-live-transcribe model pages and transcription guides publish no per-language accuracy. The guide lists supported code formats (ISO 639-1, selected ISO 639-3 and regional Chinese codes), not a language list, and says Whisper supports 98 languages with accuracy varying by language. OpenAI's own advice is to test target languages, accents, code-switching, telephony audio, and names, numbers, dates and alphanumeric strings.

AssemblyAI publishes word error rate (WER) bands for Universal-2, its 99-language model. Hindi sits in the good band (above 10% up to 25% WER). Marathi and Tamil sit in moderate (above 25% up to 50%). Assamese, Bengali, Gujarati, Kannada, Malayalam, Punjabi and Telugu sit in fair (above 50%). Its newer Universal-3.5 Pro covers 18 languages, and Hindi is the only Indian language on that list.

Google Chirp 3 lists en-IN and hi-IN as generally available, and Assamese, Bengali, Gujarati, Kannada, Malayalam, Marathi, Odia, Tamil and Telugu in Preview. Diarisation runs in BatchRecognize and Recognize, with en-IN and hi-IN on the diarisation list.

Deepgram includes Hindi in its Nova-3 Multi and Flux Multi code-switching models, and its documentation lists Bengali, Gujarati, Hindi, Kannada, Marathi, Punjabi, Tamil, Telugu and Urdu as Nova-3 monolingual options.

Azure Speech lists 13 Indian languages plus en-IN for speech to text, from Assamese and Bhojpuri to Urdu.

Sarvam covers 22 Indian languages plus Indian English with Saaras V4. It offers transcribe, translate (Indic to English), transliterate, verbatim and code-mix output modes, telephony audio at 8 kHz, key-term prompting, and batch files up to 2 hours with diarisation for up to 20 speakers.

No vendor here publishes a like-for-like Indian benchmark that the others accept, and eCorpIT has not run one. Build a 2 to 4 hour test set from your own consented recordings: Hinglish support calls, regional-language calls on 8 kHz lines, and the names, rupee amounts, PIN codes and order IDs your systems parse. Score WER and entity accuracy separately. For Indic builds on public infrastructure, see our guide to multilingual voice agents on VoicERA and BHASHINI.

Alternatives compared, with checked prices

Prices are pay-as-you-go list rates on 1 November 2026. Azure's rates come from Microsoft's retail prices API, which shows the same batch rate in Central India and East US.

Deepgram's Growth plan, from $4,000 a year in prepaid credit, advertises up to 20% off; its listed Nova-3 rates are 15% to 16% lower. Sarvam's rate drops to ₹0.38 a minute (₹0.47 with diarisation) on its ₹1,99,999-a-month Business plan.

Which option to pick

Worked cost example: 100,000 minutes of call recordings a month

Assumptions. An Indian customer-support analytics product transcribes 100,000 minutes (about 1,667 hours) of recorded Hindi and English calls each month and needs speaker labels. Storage, egress and engineering time are excluded. Sarvam's rupee price converts at ₹95 to the dollar, from the European Central Bank's 30 October 2026 rates of 1.1172 dollars and 106.084 rupees to the euro. Google's guide shows dynamic batching with Chirp 3 but its diarisation example uses standard batch, so confirm that combination. AssemblyAI warns that features vary by language and Azure's batch diarisation needs mono audio, so test speaker labels on a Hindi sample.

The cheapest OpenAI path saves $150 a month but drops speaker labels. Four vendors come in under today's bill with labels, while Sarvam costs about $21 more but publishes coverage of all 22 scheduled Indian languages. Accuracy on your own test set, not the price column, should settle it.

Live captioning changes more. Captioning 100,000 live minutes on gpt-4o-mini-transcribe at an estimated $0.003 a minute costs about $300 a month. On gpt-live-transcribe the same volume costs $1,700, and on gpt-transcribe committed turns $450.

Translation changes most. 10,000 recorded minutes translated to English on whisper-1 cost $60; through gpt-realtime-translate they cost $340 and return audio you may not need.

India and EU data residency

OpenAI's data controls page lists /v1/audio/transcriptions and /v1/audio/translations for regional storage in all listed regions, including India (in.api.openai.com), but regional processing only in the United States and Europe. Realtime transcription and translation sessions are stored and processed in the US or EU only. For DPDP planning, treat audio sent to any OpenAI transcription model as processed outside India, say so in your notice, and set retention limits. Our DPDP engineering playbook covers notices and retention.

In the EU, regional processing endpoints carry a 10% uplift for models released on or after 5 March 2026, which includes gpt-transcribe.

Microsoft states that Azure Speech stores and processes speech data in the region where you create your resource, and Central India is a listed region; Voice Live is the exception. Deepgram lists an India regional endpoint. Chirp 3 is GA only in Google's us and eu multi-regions. Sarvam's privacy policy allows transfers outside India, including to US cloud infrastructure, so confirm the processing location in its data processing terms before you send regulated audio.

Migration checklist

1. By 15 November 2026: search code and SDK wrappers for the four model IDs, audio/translations, timestamp_granularities, verbose_json, srt, vtt, diarized_json and logprobs. Add a shutdown_date check to CI.

1. Classify each call site: plain file text, file with timestamps or speakers, live captions, voice-agent input transcription, or translation.

1. By 1 December: build a consented test set of 2 to 4 hours per language mix, with reference transcripts and an entity list.

1. By 15 December: move any Azure whisper 001 deployment.

1. By 20 December: run gpt-transcribe and at least two alternatives on the test set. Record WER, entity accuracy, p95 latency and cost per hour.

1. Before 20 January 2027: move anything pinned to gpt-4o-mini-transcribe-2025-03-20.

1. By 31 January: ship the new path behind a feature flag and update DPDP notices and processor lists.

1. By 15 February: route all traffic to the new path, and keep the old one as rollback until 25 February.

            # Before (stops working on 26 Feb 2027)
client.audio.transcriptions.create(model="whisper-1", file=f, language="hi")

# After, for plain text from a recorded file
client.audio.transcriptions.create(
    model="gpt-transcribe",
    file=f,
    extra_body={"languages": ["hi", "en"], "keywords": ["UPI", "EMI"]},
)
          

How eCorpIT can help

eCorpIT is a Gurugram-based technology organisation, founded in 2021, assessed at CMMI Level 5 and MSME certified, with senior-led engineering teams working across AWS, Microsoft and Google platforms. Our AI and machine learning services team runs speech-to-text migrations like this one, from call-site inventory and Indian-language test sets to side-by-side vendor evaluation, speaker-label and subtitle rebuilds, and cutover with rollback. Talk to us at /contact-us/.

Last updated: 1 November 2026.

References

1. Deprecations — OpenAI API

1. Pricing — OpenAI API

1. Transcription — OpenAI API

1. File transcription — OpenAI API

1. Whisper model — OpenAI API

1. Model retirement schedule, Microsoft Foundry — Microsoft Learn

1. Changelog — OpenAI API

1. Create transcription, API reference — OpenAI

1. GPT-Live-Transcribe model — OpenAI API

1. Realtime transcription — OpenAI API

1. GPT-Transcribe model — OpenAI API

1. Realtime translation — OpenAI API

1. Pricing — Deepgram

1. Pricing — AssemblyAI

1. Supported languages — AssemblyAI Docs

1. Chirp 3 transcription model — Google Cloud

1. Create a batch transcription — Microsoft Learn

1. Speech to Text API — Sarvam

1. openai/whisper LICENSE (MIT) — GitHub

1. Languages and dialects — Deepgram

1. Models and languages overview — Deepgram Docs

1. Language and voice support for Azure Speech — Microsoft Learn

1. Speech-to-Text pricing — Google Cloud

1. Azure retail prices API, Azure Speech meters in Central India — Microsoft

1. API pricing — Sarvam

1. Euro foreign exchange reference rates, 30 October 2026 — European Central Bank

1. Transcribe long audio files with batch recognition — Google Cloud

1. Data controls in the OpenAI platform — OpenAI API

1. Supported regions for Azure Speech — Microsoft Learn

1. Privacy policy — Sarvam