Whisper API Pricing 2026: Every OpenAI Transcription Rate

The Whisper API (whisper-1) costs $0.006 per minute of audio, or $0.36 an hour. OpenAI's newer transcription models run from $0.003 a minute for file transcription to $0.017 a minute for live streaming, and live translation costs $0.034 a minute.

The number that matters most is the date. OpenAI removes whisper-1 from the API on February 26, 2027, along with all three GPT-4o transcription models and the /v1/audio/translations endpoint.

If you're budgeting on Whisper today, you're budgeting on a model you’ll have to replace by the end of February.

I pulled every transcription rate from OpenAI’s pricing page, including two that hide behind its “All models” toggle, and priced each migration path.

Then I lined the prices up against accuracy and speed from Cekura’s speech-to-text benchmark, so you can see which model is cheapest per minute and which is cheapest per correct transcript.

Pricing verified against OpenAI's pricing page on 28 October 2026. Verify with the vendor before you commit.

Whisper API Pricing at a Glance: Every OpenAI Transcription Model

OpenAI's API pricing page lists eight speech-to-text and speech translation models. Five bill a flat rate per minute of audio, and the three GPT-4o models bill per token, so their per-minute figures are OpenAI’s own estimates.

whisper-1 and gpt-4o-transcribe-diarize only show up on the pricing page once you click “All models”.

There are three things worth knowing before you pick a row:

• gpt-transcribe is 25% cheaper than whisper-1 and is the model OpenAI's guide now recommends for recorded speech

• gpt-4o-mini-transcribe is the cheapest OpenAI transcription rate today at about $0.003 a minute, and it shuts down with Whisper

• Live transcription costs about 3.8 times file transcription ($0.017 against gpt-transcribe's $0.0045), because you're paying for a transcript while the speaker is still talking

Whisper Shuts Down on February 26, 2027: What the Move Costs

OpenAI's deprecations page says it notified developers on August 26, 2026, and names the same two replacements for every transcription model on the list:

Most of the time, if you’re transcribing recorded audio on whisper-1, the move to gpt-transcribe means a 25% price cut. A team on gpt-4o-mini-transcribe will pay 50% more for the same minutes after the switch, since the cheapest model on the page has no cheaper successor.

What Goes Away With Whisper

The price change is the easy part. OpenAI's speech-to-text guide documents three features only on models that are shutting down:

• Word and segment timestamps run on whisper-1 through timestamp_granularities[], which captioning and video editing depend on

• Speaker labels come from gpt-4o-transcribe-diarize, and the guide says speaker labeling isn't supported in Realtime transcription sessions

• Translation to English uses /v1/audio/translations on whisper-1. Its replacement, gpt-realtime-translate, is a streaming speech-to-speech model that needs a translation session, and it costs 5.7 times as much per minute

If any of those three is in your pipeline, test gpt-transcribe against it before you budget, and leave room for a second vendor if it falls short.

File Transcription vs Live Streaming Pricing

OpenAI splits transcription into two jobs with a big price difference.

File Transcription: $0.003 to $0.006 a Minute

File models take a completed recording (up to 25 MB per request, in mp3, mp4, mpeg, mpga, m4a, wav, or webm) and return a transcript. whisper-1 and gpt-transcribe both sit here.

• gpt-transcribe ($0.0045): OpenAI's recommended model, with context prompts, keyword hints, and multiple language hints for domain terms and multilingual audio

• whisper-1 ($0.006): 98 languages, timestamps, and translation to English, until February 26

• gpt-4o-mini-transcribe (about $0.003): the cheapest rate, billed per token, until February 26

Longer recordings need splitting or compressing to fit under 25 MB, and OpenAI warns that cutting mid-sentence can lower accuracy.

Live Transcription: $0.017 a Minute

Live models take streaming audio and send transcript text while the person is still speaking. That's what a voice agent needs, since it can't reply until it has the caller's words.

• gpt-live-transcribe ($0.017): low-latency streaming with tunable latency, keyword hints, and language hints

• gpt-realtime-whisper ($0.017): streaming speech-to-text priced by audio duration, built for tuning latency against accuracy

• gpt-realtime-translate ($0.034): streaming speech-to-speech translation that returns translated audio plus a transcript

If you're transcribing recordings after the fact, streaming pays for speed you don't use. If you're feeding a voice agent, file pricing isn't an option, and the extra $0.0125 a minute will buy you a reply the caller isn't waiting on.

What the Whisper API Costs per Month

Per-minute prices are easy to misjudge at volume. This table prices each model at three monthly volumes, plus 1,000 five-minute calls:

OpenAI's pricing page lists no volume discounts on these rates, so a minute costs the same at 10,000 minutes as at a million. Google, for one, cuts its rate past 500,000 minutes a month (more on that below).

The two GPT-4o figures are estimates. They bill per token, so a long prompt (domain vocabulary, say) adds input tokens on top of the audio.

Is Whisper Free? Self-Hosted Whisper vs the API

Running the Whisper model costs nothing beyond your own hardware, and the Whisper API bills by the minute.

Can I Use OpenAI Whisper for Free?

Yes, the open-source Whisper model is free to download and run, under the MIT license. OpenAI publishes the code and weights on GitHub.

whisper-1 costs $0.006 a minute, and OpenAI's pricing page lists no free allowance. whisper-1 is the only transcription model with rate limits listed for OpenAI's Free usage tier (3 requests a minute, 200 a day), while gpt-transcribe and the live models list Free as "Not supported."

The paid tiers start at Build, which needs $5 in total credit purchases.

How to Install Whisper Locally

Whisper runs on Python with PyTorch, and it needs ffmpeg for audio decoding:

• Install the package: pip install -U openai-whisper

• Install ffmpeg: for example, brew install ffmpeg on macOS or sudo apt install ffmpeg on Ubuntu

• Transcribe a file: whisper audio.mp3 --model turbo

Model size decides the hardware you need. OpenAI's repo lists approximate VRAM and speed relative to the large model:

turbo is an optimized large-v3 with "minimal degradation in accuracy," according to the repo, and it’s the CLI’s default.

When Self-Hosting Beats the API

The question on OpenAI's community forum is whether the API is "10x more expensive" than running Whisper yourself. The answer is that it depends on how busy your GPU stays.

One developer in that thread measured a consumer GPU running an optimized Whisper build through an hour of speech in about two minutes, roughly 30 times faster than real time. Run an illustrative rented GPU at $0.40 an hour through those numbers:

• At full use, 30 hours of audio per GPU hour works out to about $0.013 per audio hour, against $0.27 on gpt-transcribe. The thread’s 30x figure comes from the same math against whisper-1’s $0.36.

• Left running all month, the same GPU costs about $292, busy or idle. You'd need about 1,080 hours of audio a month (65,000 minutes) before that beats gpt-transcribe, and about 810 hours to beat whisper-1.

Below that volume, the API is cheaper, and it's cheaper still once you count what self-hosting adds, like queuing, retries, GPU autoscaling, model updates, and an engineer who owns it.

Above it, self-hosting wins on unit price, with one more catch for voice agents. The open-source Whisper models transcribe files, so a live agent needs a streaming layer you build yourself.

Whisper API Pricing vs Other Speech-to-Text APIs

Here’s how OpenAI compares with other speech-to-text APIs at each vendor’s published pay-as-you-go rate:

For recorded audio, OpenAI sits in the middle of the pack. Google’s dynamic batch, Azure’s batch rate, Deepgram’s Nova-3, and AssemblyAI’s Universal-2 all come in under gpt-transcribe’s $0.0045.

For streaming, OpenAI's $0.017 is the top of the range, more than three times Deepgram's promotional Nova-3 rate.

Azure also hosts Whisper itself. Don’t plan on it as a way around the shutdown, since Microsoft’s retirement schedule takes Azure OpenAI’s whisper model offline on December 15, 2026, more than two months before OpenAI’s own date.

Per-minute price is one axis, though, and a cheap rate stops looking cheap once someone has to fix the transcript.

Price per Correct Transcript: What Cekura's STT Benchmark Shows

Cekura's speech-to-text benchmark (last updated September 16, 2026, prices checked October 8) streams the same audio through 19 STT models and scores word error rate (WER) and how fast the final transcript arrives.

It uses two datasets: 1,000 short public clips from Pipecat's STT dataset, and eight recordings from four real phone conversations licensed from Ocular.

Three OpenAI models are in it. whisper-1, gpt-transcribe, and gpt-live-transcribe aren’t, so there’s no benchmark figure for them.

To make those percentages concrete, take an illustrative hour of phone audio at 150 words a minute, about 9,000 words:

• GPT Realtime Whisper: about 320 wrong, missing, or extra words, at $1.02

• GPT-4o Mini Transcribe: about 400, at $0.18

• GPT-4o Transcribe: about 1,120, at $0.36

So GPT-4o Transcribe costs twice what the mini does and gets nearly three times as many words wrong on real calls. GPT Realtime Whisper is the most accurate OpenAI model in the benchmark on phone audio, and the most expensive per hour of anything in this table.

Speed counts for more than it looks on a voice agent. "Final text" is how long after the caller stops talking the full transcript arrives, and the agent can't start its reply before then.

OpenAI's three models all take more than half a second at the median. The benchmark's fastest, Soniox STT-RT v5, delivers final text in 38 milliseconds. The GPT Realtime Whisper benchmark page has its P90 and P95 timings.

Accuracy also shifts with who's talking. Our breakdown of accent and dialect testing for speech-to-text covers how to check a model on your own callers before you trust a leaderboard number.

Which OpenAI Transcription Model Should You Use?

Choose gpt-transcribe if you:

• Transcribe recorded calls, meetings, or uploads after the fact

• Are on whisper-1 today and want a cheaper, supported replacement before February 26

• Need domain terms handled through context prompts or keyword hints

Choose gpt-4o-mini-transcribe if you:

• Want the lowest OpenAI rate for recorded audio right now

• Can plan a second migration before February 26, when it shuts down

Choose gpt-realtime-whisper or gpt-live-transcribe if you:

• Feed a voice agent that has to reply mid-conversation

• Need the best phone-audio accuracy OpenAI has benchmarked, which is gpt-realtime-whisper

Look outside OpenAI if you:

• Run high volume, where Google and others offer tiers or batch rates below OpenAI's flat price

• Need streaming at a lower rate or a faster final transcript

• Depend on word timestamps, speaker labels, or translation to English, and gpt-transcribe falls short in your tests

Test the Transcript Before You Trust the Price

Whisper API pricing tells you what a minute of audio costs, and the error rate on your callers’ accents, product names, and noisy lines decides what that minute is worth.

The answer for your agent comes from running your own audio through each candidate before February 26 forces the switch. Cekura runs that testing on full simulated calls, so you see what a transcription error does to the conversation as well as the WER.

With Cekura, teams can:

• Simulate callers with different accents, speaking speeds, and background noise, and see where each model's transcript breaks the conversation

• Compare a whisper-1 agent against the same agent on gpt-transcribe or another STT vendor, scenario by scenario, before the cutover

• Re-run your saved test scenarios after a model change to catch regressions before they reach callers

• Monitor live calls for misheard inputs, drop-offs, and missed workflow steps once the new model is in production

Native integrations cover Retell, Vapi, ElevenLabs, LiveKit, Pipecat, and Bland, and Cekura is SOC 2-, HIPAA-, and GDPR-compliant for transcript redaction, role-based access, and audit trails.

Book a demo to test your voice agent's transcription before the Whisper API shutdown.

Frequently Asked Questions

Is the Whisper API Free?

No, the Whisper API isn't free: whisper-1 costs $0.006 per minute of audio. The open-source Whisper model is free to run yourself under the MIT license, so you pay only for your own hardware. OpenAI's pricing page lists no free minutes for any transcription model.

When Is Whisper Being Shut Down?

OpenAI shuts down whisper-1 on February 26, 2027. gpt-4o-transcribe, gpt-4o-mini-transcribe, gpt-4o-transcribe-diarize, and the /v1/audio/translations endpoint shut down the same day.

OpenAI recommends gpt-transcribe for recorded audio and gpt-live-transcribe for live audio, and gpt-realtime-translate replaces the translations endpoint.

What Is the Cheapest OpenAI Transcription Model?

gpt-4o-mini-transcribe is the cheapest OpenAI transcription model, at about $0.003 per minute. It shuts down on February 26, 2027. Among the models that stay, gpt-transcribe is cheapest at $0.0045 per minute, 25% below whisper-1's $0.006.

Is There a Whisper API Pricing Calculator?

No, OpenAI doesn't publish a Whisper API pricing calculator, and you don't need one for most models.

Multiply your audio minutes by the per-minute rate. 10,000 minutes on gpt-transcribe costs $45, and on whisper-1, $60. The GPT-4o transcription models bill per token, so their per-minute rates are OpenAI's estimates.

How Do I Get a Whisper API Key?

You get a Whisper API key by creating an API key in your OpenAI platform account, the same key used for every OpenAI model. whisper-1 lists low rate limits on the Free usage tier (3 requests a minute), and the newer transcription models need a paid tier, which starts at $5 in total credit purchases.

How Long Can an Audio File Be on the Whisper API?

An audio file on the Whisper API can be up to 25 MB per request. OpenAI's guide says to compress longer recordings or split them into chunks of 25 MB or less, and to avoid splitting mid-sentence, which can reduce accuracy.