If your product has a Whisper call in it, that call has about four months left.

OpenAI deprecated whisper-1 on 26 August 2026 and removes it from the API on 26 February 2027, along with three other transcription models [1]. Nothing broke. It is being retired, which turns a line of code nobody has opened since the integration shipped into a dated engineering task.

For teams running clinical audio, that forced rewrite is worth more than it costs. It is the only moment anyone re-examines a transcription choice, and most teams spend it porting the same assumptions to a new endpoint.

Key Takeaway: OpenAI deprecated whisper-1 on 26 August 2026 and removes it from the API on 26 February 2027, along with gpt-4o-transcribe, gpt-4o-mini-transcribe and gpt-4o-transcribe-diarize. Teams running clinical audio are migrating on a deadline, not a preference. OpenAI's recommended replacements are gpt-transcribe for recorded audio and gpt-live-transcribe for live audio. For clinical audio the harder question is not price or word error rate but fabrication. Peer-reviewed work found that roughly 1% of Whisper transcriptions contained an entire phrase that was never spoken, and 38% of those hallucinations carried an explicit harm, and OpenAI's own model card recommends against use in high-risk decision-making contexts. Clinical transcription APIs built for this workload, Sully among them, let you boost your own medication and specialty terms in the API call and return the finished note rather than a transcript you still have to process.

Why Teams Running Clinical Audio Are Suddenly Shopping for a Whisper API Alternative

Nobody woke up unhappy with Whisper. It was cheap, it was good enough, and it was already in the codebase. The shutdown is what moved it from the backlog to the roadmap.

OpenAI notified developers on 26 August 2026. Removal is 26 February 2027 [1].

The second point matters more than the first. If you are rewriting the integration anyway, the cost of moving to a different vendor is now close to the cost of staying with OpenAI. Teams almost never get a free look at a vendor decision. This is one.

The Four Models Leaving the API

Four transcription models go at the same time: whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe and gpt-4o-transcribe-diarize [1].

The /v1/audio/translationsendpoint goes with them, replaced by v1/realtime/translations [1].

If you are running the open-source Whisper weights on your own hardware, the deadline does not touch you. This is the hosted API only. The clinical gaps still do: self-hosting removes the shutdown date and leaves you owning the GPU bill, the diarization bolt-on and the missing custom vocabulary [2].

What gpt-transcribe and gpt-live-transcribe Replace Them With

OpenAI recommends gpt-transcribe for recorded audio and gpt-live-transcribe for live audio [1]. The transcription guide names gpt-transcribe as the default for transcribing recorded speech in its original language [3].

One genuine improvement comes with the change. whisper-1never supported response streaming on completed files; the newer file-transcription path does [3]. That is a real gain and worth saying before arguing about anything else.

It is also the only one of this article's clinical checks that the upgrade answers.

What Whisper API Pricing Looks Like Now and What the Replacement Costs

The number everyone searches for is $0.006 per minute, or $0.36 an hour of audio [4]. That price is attached to a model with four months left on it.

The four removed models' prices come from OpenAI's published pricing [4]. The two replacement prices come from Cartesia's pricing round-up [5] and were not on OpenAI's own pricing table at the time of writing, so check them before you budget.

Read the bottom two rows together. Recorded audio gets cheaper. Live audio costs roughly three times what the batch model cost, which is the number to put in front of anyone who assumed real-time transcription was a configuration change.

Neither figure covers what a clinical pipeline still has to build on top of the transcript, which is the rest of this article.

Where the Per-Minute Price Stops Being the Whole Bill

The hosted API caps uploads at 25 MB, across mp3, mp4, mpeg, mpga, m4a, wav and webm [3].

Clinical audio runs long. A 40-minute encounter at any reasonable bitrate clears that cap, so somebody writes chunking logic. OpenAI's own guidance warns that splitting mid-sentence removes context and reduces accuracy [3], which means the chunking has to be smart enough to find a pause.

That is engineering time, not an API line item, and it never appears in a per-minute comparison.

The Speaker Diarization Gap Nobody Has Priced In

Here is the consequence of the shutdown that none of the pricing pages have worked through.

As OpenAI's transcription guide stands today, speaker diarization runs only through gpt-4o-transcribe-diarize, using the diarized_jsonresponse format, and it is not available in realtime transcription sessions at all [3].

That model is one of the four being removed in February [1].

For a clinical product that has to tell the clinician from the patient, diarization is not a nice-to-have. It is how a transcript becomes a note. OpenAI may ship diarization on gpt-transcribebefore the deadline, so check the documentation on the day you plan the migration. As it reads now, that path closes.

The Failure Mode a Word Error Rate Cannot Score

The reason Whisper is a poor fit for clinical audio has nothing to do with how many words it gets wrong.

What the Research Actually Measured

Koenecke and colleagues, presented at the ACM Conference on Fairness, Accountability, and Transparency in 2024, evaluated Whisper across more than 13,000 speech clips [6][7].

Roughly 1% of transcriptions contained an entire phrase or sentence that did not exist anywhere in the audio. Of that hallucinated content, 38% contained an explicit harm: perpetuating violence, inventing personal details, or implying an authority that was never there [6].

Read the second number carefully. The 38% is a share of the hallucinations, not of the transcriptions. In round terms, fewer than four in a thousand transcriptions carried an explicitly harmful fabrication.

Two caveats belong in the same breath. The evaluation ran on a 2023 release of Whisper, and Cornell reports that OpenAI has improved the model since and the hallucination rate has come down [7].

The field reports are messier and harder to forget. One machine learning engineer found hallucinations in about half of the more than 100 hours of Whisper transcriptions he analyzed. A University of Michigan researcher found them in eight of every ten transcriptions he inspected. One clinical transcript produced a medication that does not exist, "hyperactivated antibiotics" [8].

Those are observations, not rates. They are also the kind of thing that ends up in a deposition.

Why a Reference Transcript Cannot Catch It

Word error rate counts substitutions, insertions and deletions against a reference transcript. A sentence invented during a silent gap scores as a run of insertions.

So a model that fabricates a whole clause and a model that mishears a single drug name can land in the same neighborhood of a benchmark number, while meaning completely different things for a patient chart. If you want the longer version of that argument, it is the same problem as scoring a vendor on aggregate word error rate instead of on your own term list.

OpenAI says the same thing in its own documentation. The Whisper model card states that "the predictions may include texts that are not actually spoken in the audio input," and attributes it to the model combining next-word prediction with transcription [9].

That behavior is a property of the design. It is not a bug that a version bump removes.

Why Clinical Audio Makes It Worse

The research found hallucinations occurred more often for speakers with aphasia, who have longer non-vocal stretches, than for the control group [6].

The mechanism it identified is silence. Clinical audio is full of silence: a clinician typing, a patient undressing, a pause while someone reads a chart, the thirty seconds a blood pressure cuff takes to inflate.

The study measured aphasia specifically, so this is a mechanism worth testing on your own recordings rather than a finding to assume. It is also the most obvious place to start testing.

OpenAI's own line closes the section better than ours would. The Whisper model card says: "We recommend against use in high-risk domains like decision-making contexts" [9].

Two Things Every Whisper Alternatives Guide Gets Wrong About Clinical Use

Seven of the top-ranking pages for this topic were checked while writing this article. They compare the same four things: accuracy, streaming, latency and pricing [10].

Not one of them asks what the audio is of. Two of the most visible get the clinical question actively wrong.

OpenAI Will Sign a BAA, and That Is Not the Same as Zero Retention

The claim in circulation is that OpenAI does not offer a HIPAA Business Associate Agreement, so you cannot send protected health information through a general-purpose model. It appears in pricing guides that Google's AI answer currently cites.

It is wrong. OpenAI's enterprise privacy page states: "We are able to sign Business Associate Agreements (BAA) in support of customers' compliance with the Health Insurance Portability and Accountability Act (HIPAA). Please reach out if you require a BAA." [11]

The real gap is a different one, and it is the question a product manager actually has to take to legal. OpenAI may retain API inputs and outputs for up to 30 days to provide the service and identify abuse, and zero data retention is available only "for eligible endpoints if you have a qualifying use-case" [11].

Availability on request is not the same as a default, and the retention posture underneath the agreement is the thing worth comparing. Sully's APIs run Zero Data Retention, with nothing stored after processing and no training on customer data [12], under ISO 27001, HIPAA, HITRUST and SOC 2 Type II [12][13].

Word Error Rate Is Not Why Whisper Fails on Clinical Audio

The second thing the guides get wrong is subtler, because it is a framing error rather than a factual one.

They treat hallucination as a point on the accuracy axis, somewhere below a good word error rate. It is not on that axis. An invented sentence is a different category of output from a misheard one, and the instrument most of these comparisons use cannot tell them apart.

Which changes what you should be testing in a replacement.

What to Check Before You Move Clinical Audio to Another Transcription API

1. Whether You Can Boost Your Own Term List in the API Call

Ask whether the vendor lets you pass your own medication, device and specialty terms at request time.

Whisper's managed API offers no custom vocabulary at all [2], which is why so many teams built a correction layer downstream. Sully's streaming transcription boosts the words that matter, such as drug names, in the call itself [12].

2. What Happens to the Audio and the Transcript After the Call Returns

Not "are they HIPAA compliant." Three specific questions: is a BAA available and what does it cover, what the default retention window is, and whether zero data retention is available on the exact endpoint you are calling.

The third one is the one teams skip, and it is the one that differs most between vendors [11].

3. Whether Diarization and Streaming Come From One Vendor or Three

The Whisper ecosystem solved diarization and streaming with bolt-ons: WhisperX, faster-whisper, pyannote.

Every one of those is another dependency to upgrade, another place audio sits on disk, and another thing your BAA has to cover. It is the same assembly problem Amazon Transcribe Medical leaves clinical teams with, in a different ecosystem.

Ask whether both ship natively.

4. What the Vendor Hands Back After the Transcript

A transcript is not a note, and a note is not an EHR write.

Ask how many steps past the transcript the vendor goes. Sully returns the finished note from a transcript plus your own template, and validated fields against your own JSON Schema for the write-back [12]. This is the same gap that shows up when you compare Deepgram alternatives for clinical audio: the speech vendors stop at the transcript.

5. How Much Migration Window You Actually Have

Count backwards from 26 February 2027 [1] through evaluation, integration, clinical validation and whatever parallel-run period your team needs before it trusts a new transcript.

Most of that is not engineering time. Put the number on the roadmap before someone else discovers it in January.

How Sully Handles the Same Workload

Sully's APIs were built for this workload rather than adapted to it.

Transcription takes a recorded visit up to 24 hours long across 30+ languages. Streaming transcription runs during the encounter and lets you boost the words that matter, such as drug names. Note generation takes a transcript and your own template and returns the finished note. Text to JSON takes a note or transcript with your JSON Schema and returns validated fields, ready for the EHR. Sully M1 is the general clinical model, called like any chat model, for the things the task APIs do not cover [12].

The platform runs 100k+ notes per month across 40+ specialties, at 99.96% uptime and 11.23s p50 latency [12].

Tebra's AI Scribe is built on Sully [12]. That is the proof worth more than any benchmark on this page: another health-tech company shipped a product on these APIs.

Now the part that belongs in an article arguing against vendor accuracy claims. Sully publishes no transcription word error rate and no hallucination rate. What we publish is 92.2% clinical note accuracy, and that figure measures the finished note, not the transcript underneath it [12]. It is not comparable to the numbers in this article, and it would be dishonest to put it beside them.

What we will say is that the checks above are the ones we built for. Boost your own terms in the call. Sign a BAA by default and retain nothing. Return something further down the pipeline than a transcript.

Run the test instead. Get an API key at dashboard.api.sully.ai, point it at 20 encounters from your own product, and check the medication names against the audio. Change the base URL and the model name. There is no demo to sit through [12].

FAQ

Q: Is the Whisper API free? No. OpenAI's hosted Whisper API is priced at $0.006 per minute of audio [4]. The Whisper model weights are open source and free to download, so self-hosting removes the per-minute charge and replaces it with GPU, storage and DevOps costs that only pay off at high volume. From 26 February 2027 the hosted whisper-1endpoint stops serving traffic entirely [1], so the free-versus-paid question is being replaced by a migration question.

Q: How much does the Whisper API cost per minute? $0.006 per minute on OpenAI's published pricing, which works out to $0.36 for an hour of audio [4]. The replacement model OpenAI recommends for recorded audio, gpt-transcribe, is listed at about $0.0045 per minute in third-party pricing round-ups [5]. Check OpenAI's own pricing page before you budget, because the replacement models were not yet listed on it at the time of writing.

Q: Does the Whisper API support speaker diarization? Not on whisper-1. As OpenAI's transcription guide stands today, diarization is available only through gpt-4o-transcribe-diarize using the diarized_jsonresponse format, and it is not supported in realtime transcription sessions [3]. That model is one of the four being removed on 26 February 2027 [1], so any clinical product relying on it needs a replacement plan. Teams that need diarization today usually bolt on a separate library such as pyannote, which puts another dependency inside the boundary your BAA has to cover.

Q: Does OpenAI offer HIPAA compliance for the Whisper API? OpenAI states it is able to sign Business Associate Agreements for the API Platform in support of customers' HIPAA compliance, on request [11]. Several widely-read pricing guides claim the opposite, and they are wrong. The better question is about default data handling: OpenAI may retain API inputs and outputs for up to 30 days, and zero data retention applies only to eligible endpoints with a qualifying use case [11]. Ask every vendor the same pair of questions, including this one. Sully's answer is Zero Data Retention by default, with nothing stored after processing and no training on customer data [12].

Q: What is the best Whisper API alternative for medical audio? There is no single answer, because the alternatives split by what you need past the transcript. If you only need a drop-in transcript, the general speech APIs are close substitutes for each other. If you are building a clinical product, the checks that decide it are whether you can boost your own medication and specialty terms in the API call, whether diarization and streaming come from one vendor, what the default retention posture is, and how far past the transcript the vendor goes. Sully's APIs cover transcription, streaming with word boosting, note generation against your own template, and schema-validated fields for the EHR [12].

Sources

[1] OpenAI — Deprecations [2] Gladia — What is OpenAI Whisper? [3] OpenAI — File transcription [4] OpenAI — API Pricing [5] Cartesia — OpenAI Whisper API pricing, model by model[6] Koenecke, Choi, Mei, Schellmann and Sloane — Careless Whisper: Speech-to-Text Hallucination Harms [7] Cornell Chronicle — AI speech-to-text can hallucinate violent language[8] Associated Press — Researchers say an AI-powered transcription tool used in hospitals invents things no one ever said [9] OpenAI — Whisper model card [10] AssemblyAI — 7 best Whisper alternatives for speech-to-text[11] OpenAI — Enterprise privacy at OpenAI [12] Sully.ai — Sully APIs [13] Sully.ai — Trust Center