OpenAI Whisper API Pricing in 2026: Real Billed Costs and the 2027 Shutdown

OpenAI Whisper API pricing is $0.006 per minute, or $0.36 per hour of audio. Two things make that number less simple than it looks. OpenAI now lists a cheaper file model, gpt-transcribe, at $0.0045 a minute, and whisper-1 shuts down on February 26, 2027, together with the endpoint that translates speech into English.

We sent the same four sets of audio to Whisper and four other speech-to-text models and recorded what each task billed: 39 completed tasks, 45 minutes of audio per model, $0.72 in total. Below are the price tables, how short files are billed, word error rates, the names each model got wrong, and the features you lose when Whisper goes away. If the shutdown notice brought you here, you need two answers: what Whisper costs until then, and which model to move to. The test is set up so you can rerun it on your own audio before you pick.

OpenAI Whisper API Pricing at a Glance: $0.006 a Minute

Source: OpenAI API pricing, Transcription section, and the deprecations page. OpenAI labels these per-minute figures as estimated cost. The 4o transcription models are billed by audio tokens underneath, at $2.50 per 1M input tokens for gpt-4o-transcribe and $1.25 for the mini model. For OpenAI's wider price list, including other API products, see OpenAI pricing plans.

Whisper API cost calculator: price per hour

Multiply hours of audio by the hourly rate. Rate limits and file size do not change the price.

Grok Voice ASR is the lowest list price among the five models we tested. Whisper costs 3.6 times as much for the same hour of audio.

What Whisper Actually Bills: Short Files, File Limits, and Rate Limits

How short files are billed

OpenAI does not publish a rounding rule for its transcription API, so we measured one where we could. We cut four clips from the same LibriVox recording, 5, 31, 61, and 119 seconds long, and sent each once to every model on Modellix.

Model on Modellix	5 s clip	31 s clip	61 s clip	119 s clip
Whisper	$0.0005	$0.0031	$0.0061	$0.0119
MAI Transcribe 1.5	$0.0005	$0.0031	$0.0061	$0.0119
Gemini 3.5 Transcribe	$0.0002	$0.0015	$0.0030	$0.0059
Fun-ASR	$0.0001	$0.0009	$0.0019	$0.0036
Grok Voice ASR	$0.0001	$0.0008	$0.0016	$0.0033

None of the five models charged a full minute for a short clip. On Modellix, Whisper and MAI Transcribe 1.5 billed $0.0001 per second, so the 5 second clip cost $0.0005 instead of the $0.006 a one-minute minimum would charge. Gemini 3.5 Transcribe and Grok Voice ASR billed per second and cut the amount to four decimal places. These results describe Modellix billing only; they say nothing about how OpenAI bills its own endpoint.

Fun-ASR billed less than the file length on every clip. In our test runs, its charge matched the total length of the sentences it detected: 685.7 billed seconds against 685.6 seconds of detected speech on the 802 second reading, and 140 of 180 seconds on the podcast clip with its music intro. Silence and music were not billed in those tasks. Modellix has not published a formal rule that Fun-ASR bills only detected speech; the figures above are the observed charges from this test.

The 25 MB file limit

Whisper accepts files up to 25 MB in mp3, mp4, mpeg, mpga, m4a, wav, or webm (OpenAI speech-to-text guide). A 64 kbps mono mp3 fits about 50 minutes under that cap. Our 25.5 minute test file was 12.2 MB and went through in one task. For longer recordings, split the file and send the pieces; you pay for the same total minutes, but you need to stitch the text back together yourself.

Rate limits by usage tier

The whisper-1 model page lists 3 requests per minute and 200 per day on the Free tier, then 2,500, 7,500, and 10,000 requests per minute on the Build, Launch, and Grow tiers. Rate limits cap throughput; they do not change the price per minute.

whisper-1 Shuts Down on February 26, 2027: What to Migrate

OpenAI announced the deprecation on August 26, 2026. On February 26, 2027 these stop working: whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe, gpt-4o-transcribe-diarize, and the /v1/audio/translations endpoint. OpenAI's recommended replacement for recorded audio is gpt-transcribe, and gpt-live-transcribe for live audio.

The swap to gpt-transcribe

For plain transcription, the change is the model name, and the price drops 25%, from $0.006 to $0.0045 a minute:

from openai import OpenAI

client = OpenAI()  # reads OPENAI_API_KEY from the environment
with open("meeting.mp3", "rb") as audio:
    result = client.audio.transcriptions.create(
        model="gpt-transcribe",  # was "whisper-1"
        file=audio,
    )
print(result.text)

The request still goes to POST /v1/audio/transcriptions. OpenAI's migration guide says not to assume that text, verbose_json, SRT, or VTT response formats stay valid on gpt-transcribe.

What has no OpenAI replacement yet

The OpenAI speech-to-text guide now says its examples for word timestamps, speaker labels, and subtitle formats use deprecated models. If your product depends on any of these, check where they will come from before 2027.

Source: OpenAI's migration guide, section 8. Retain whisper-1 for native SRT or VTT and for word or segment timestamps. Speaker labels on OpenAI's API use gpt-4o-transcribe-diarize, not gpt-transcribe.

Compare replacements on one key before you switch

gpt-transcribe is not the only replacement, and the lowest priced model in our test lists at $0.10 an hour against its $0.27. Before you commit, run your own audio through several models and compare transcripts and bills side by side. On Modellix, every model in this article runs on one API key with the same submit and poll flow, so the comparison is a short script:

1. Pick one audio URL from your real traffic, ideally a file with names, numbers, or two speakers.
2. Submit it to each model. The model path changes, and for two models so does the audio field: Whisper, MAI Transcribe 1.5, and Grok Voice ASR take url, Gemini 3.5 Transcribe takes audio, and Fun-ASR takes file_url.
3. Poll each task_id, then compare the transcript and the billing.amount of each task.
4. Point production at the model that wins on your audio. Auth, polling, and result handling stay the same.

AUDIO="https://example.com/meeting.mp3"

for m in openai/whisper-1 microsoft/mai-transcribe-1.5 xai/grok-voice-asr; do
  curl -s -X POST "https://api.modellix.ai/api/v1/$m" \
    -H "Authorization: Bearer $MODELLIX_API_KEY" \
    -H "Content-Type: application/json" \
    -d "{\"url\": \"$AUDIO\"}"
done

curl -s -X POST https://api.modellix.ai/api/v1/google/gemini-3.5-transcribe \
  -H "Authorization: Bearer $MODELLIX_API_KEY" \
  -H "Content-Type: application/json" \
  -d "{\"audio\": \"$AUDIO\"}"

curl -s -X POST https://api.modellix.ai/api/v1/alibaba/fun-asr \
  -H "Authorization: Bearer $MODELLIX_API_KEY" \
  -H "Content-Type: application/json" \
  -d "{\"file_url\": \"$AUDIO\"}"

Each call returns a task_id; poll it with GET /api/v1/tasks/<task_id> as in the REST example further down. One limit to plan for: if you depend on SRT or VTT files, none of the four alternatives returns subtitle files directly, so you would build captions from word timestamps yourself.

Same Audio, 5 Transcription APIs: What Each One Billed

We used public-domain audio so anyone can rerun the test: two LibriVox readings with Project Gutenberg texts as the reference, and the first 3 minutes of NASA's Houston We Have a Podcast episode 417 for names and numbers. Every model got each file once. WER counts substituted, missing, and extra words after lowercasing, removing punctuation, and spelling out numbers.

Model	List price	Billed per hour	WER, 13.4 min	WER, 25.5 min
Whisper	$0.360/hr	$0.360	1.72%	1.44%
MAI Transcribe 1.5	$0.360/hr	$0.360	1.15%	1.19%
Gemini 3.5 Transcribe	$0.180/hr	list price used¹	0.33%	0.71%
Fun-ASR	$0.126/hr	$0.107²	1.58%	1.11%
Grok Voice ASR	$0.100/hr	$0.100	1.15%	2.04%

¹ Gemini 3.5 Transcribe billed correctly on clips up to 2 minutes. Its five longer tasks finished with a $0 charge, which is a platform billing error, not a price. We use its $0.18 list price everywhere on this page. We have not retested those longer files after a possible billing fix. ² Fun-ASR billed detected speech, not file length (see short files above).

Whisper billed exactly its list price, $0.36 an hour, and was the slowest: 76 seconds to return the 13.4 minute reading and 72 seconds for the 25.5 minute one. Grok Voice ASR returned the same files in 6 and 8 seconds.

Names, numbers, and speakers in a podcast clip

The NASA clip has a host, a guest, a promo read, launch audio, and 20 terms we could check: crew names, NASA programs, dates, mission lengths, and a URL.

Model	Terms caught	What it got wrong	Speakers found
Whisper	18/20	Koch as "Cook", host name as "Niloufar"	No diarization
MAI Transcribe 1.5	18/20	Koch as "Cook", host name as "Niloufer"	No diarization
Gemini 3.5 Transcribe	20/20	None of the 20	4, no timestamps returned
Fun-ASR	17/20	Host name as "nil for randy", Koch as "cook", names in lowercase	3, six turns with times
Grok Voice ASR	18/20	Host name as "Nilfa Ramsey", Reid as "Reed"	2, labels on each word

Press the sound button on any card to hear the same 26 seconds of the NASA clip: the host introducing herself (0:36 to 0:48) and the crew list (1:48 to 2:01). NASA spells the names Nilufar Ramji and Christina Koch. Each card shows how that model wrote them in its transcript of the full 3 minute clip, from our test run. MAI Transcribe 1.5 is left out here because it spelled the crew name the same way as Whisper (Christina Cook); its result is in the table above. Try any of them in the playground.

Whisper (whisper-1)

Host: “Niloufar Ramji” · Crew: “Christina Cook”

Copy PromptTry it

Gemini 3.5 Transcribe

Host: “Nilufar Ramji” · Crew: “Christina Koch”

Copy PromptTry it

Fun-ASR

Host: “nil for randy” · Crew: “christina cook”

Copy PromptTry it

Grok Voice ASR

Host: “Nilfa Ramsey” · Crew: “Christina Koch”

Copy PromptTry it

Open Whisper in the Playground

A public-domain presidential speech

The same card pattern works for official speeches that are in the public domain. This clip is 19.4 seconds from President Kennedy's inaugural address of January 20, 1961, starting at "And so, my fellow Americans." The recording is public domain: it is published on Wikimedia Commons as public domain, credited to the John F. Kennedy Presidential Library and Museum. The official wording on the card follows the American Presidency Project transcript of the inaugural address. We have not run this clip through the five models on Modellix; that transcription is pending.

Press the sound button to hear the excerpt. The card shows the official transcript, not a model output. Audio: JFK Library / Wikimedia Commons, public domain.

JFK inaugural address

Official transcript. Model transcription pending.

Copy PromptTry it

Open Whisper in the Playground

Where the cheaper models slipped

On the 25.5 minute reading, Grok Voice ASR had the highest WER at 2.04%; it turned "rubicund" into "kund" and dropped the "tut, tut" in one line of dialogue. Fun-ASR wrote "Morris" as "Maurice" and "paw" as "poor" in the story whose title is The Monkey's Paw. Whisper made the same "paw" error twice. On the podcast, three of five models heard astronaut Christina Koch as "Christina Cook". Whisper was not the most accurate model in this test: MAI Transcribe 1.5, Gemini 3.5 Transcribe, and Fun-ASR all had a lower WER on both readings.

Compare the five models on one key →

Whisper API vs Other Speech-to-Text APIs: Price per Hour and How to Choose

Provider	Model	Per minute	Per hour
OpenAI	whisper-1	$0.006	$0.36
OpenAI	gpt-transcribe	$0.0045	$0.27
Google Cloud	Speech-to-Text V2, dynamic batch	$0.003	$0.18
AssemblyAI	Universal-2	$0.0025	$0.15
AssemblyAI	Universal-3.5 Pro	$0.0035	$0.21
ElevenLabs	Scribe v2	$0.0037	$0.22
Deepgram	Nova-3 Monolingual, pre-recorded (pay as you go)	$0.0043	$0.26
Modellix	Fun-ASR	$0.0021	$0.126
Modellix	Grok Voice ASR	$0.0017	$0.100
Modellix	Gemini 3.5 Transcribe	$0.0030	$0.180

List prices from each provider's pricing page, before volume discounts. Rows marked Modellix are prices on our own platform, from the model pages. The Deepgram row is Nova-3 Monolingual pre-recorded on pay as you go, from the Pre-recorded table on Deepgram pricing.

Pick a Whisper replacement by workload

Each row links to a model page on the same key. Run your own file there, read the transcript, and check what the task billed before you move production traffic.

If your pipeline also turns text back into speech, for example dubbing a transcript or reading it back in a cloned voice, see our MiniMax voice clone guide; it runs on the same API key.

Is self-hosting open-source Whisper cheaper?

The Whisper model weights are open source under the MIT license (openai/whisper on GitHub), so running it yourself has no per-minute fee. You pay for the GPU or CPU time, storage, and the work of keeping a transcription service up. Whether that beats $0.10 to $0.36 an hour depends on how many hours you transcribe and whether the hardware sits idle between jobs. The self-hosted model also keeps working after February 2027, since the shutdown applies to OpenAI's hosted API.

Run whisper-1 and its alternatives on one API key

Whisper on Modellix costs the same $0.36 an hour as OpenAI's list price, so your current workload keeps running while you test replacements next to it. Each model on this page is one path on the same key, which is how we ran this test. These are the ways to call them.

Playground

No code. Open a model page such as Whisper, add an audio file or URL, and the transcript and the task cost show up in your browser.

Coding agent, no install

Paste one sentence into Claude Code, Codex, or Cursor and let it read the docs:

My Modellix API key is mdlx-••••••••••••••••••••••••. Read https://docs.modellix.ai/get-started/index.md and https://docs.modellix.ai/openai/whisper-1.md , then transcribe https://example.com/meeting.mp3 with openai/whisper-1, response_format verbose_json, then run the same URL on xai/grok-voice-asr, google/gemini-3.5-transcribe, alibaba/fun-asr, and microsoft/mai-transcribe-1.5, using each model's own audio field from the docs. Save every transcript to ./transcripts and give me a table of what each task billed.

Swap in your own key from the console, and only share it with agents you trust.

Agent Skill

Install the skill once and your agent reads each model's schema and fills in the request fields itself. Setup is in the Skill guide.

npx skills add https://github.com/Modellix/modellix-plugin --skill modellix-design

Plugin

If your host has a plugin marketplace (Claude Code, Codex, Cursor), the plugin adds the Modellix skills and the Media Generation MCP. In Claude Code, run /plugin marketplace add Modellix/modellix-plugin, then /plugin install modellix@modellix. Other hosts are in the Plugin guide.

CLI

For scripts and cron jobs. One command submits the task, waits, and prints the result. See the CLI guide.

npm install --global modellix-cli
modellix-cli model run --model-slug openai/whisper-1 --body-file request.json --wait

REST API

Submit the audio URL, keep the task_id, and poll until the status is success:

curl -X POST https://api.modellix.ai/api/v1/openai/whisper-1 \
  -H "Authorization: Bearer $MODELLIX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"url": "https://example.com/meeting.mp3", "response_format": "verbose_json"}'

curl "https://api.modellix.ai/api/v1/tasks/<task_id>" \
  -H "Authorization: Bearer $MODELLIX_API_KEY"

The finished task links to a transcript JSON with text, channels, and duration_ms, plus the amount billed. The audio field is named differently per model (url for Whisper and Grok, audio for Gemini, file_url for Fun-ASR), so check each schema in the API guide.

OpenAI shuts down hosted whisper-1 on February 26, 2027. Modellix has not published whether the openai/whisper-1 listing will be retired, kept, or rerouted after that date, so plan the switch well before then. If you already call Whisper through Modellix, moving to another model on this page is a path change, plus a renamed audio field for Gemini 3.5 Transcribe and Fun-ASR.

Speech-to-text models on the same key

WhisperGrok Voice ASRGemini 3.5 TranscribeFun-ASRMAI Transcribe 1.5

OpenAI Whisper API Pricing FAQ (2026)

How much does it cost to use Whisper AI?

The OpenAI Whisper API costs $0.006 per minute of audio, which is $0.36 an hour or $360 for 1,000 hours. The same model costs the same on Modellix. The open-source Whisper model is free to download and run on your own hardware.

Is Whisper free?

The open-source model is free: the code and weights are MIT licensed and run offline on your own machine. The OpenAI Whisper API is not free; the free usage tier only limits how many requests you can send (3 a minute, 200 a day) and you still pay per minute of audio.

How is OpenAI transcription pricing calculated?

By the length of the audio. OpenAI lists an estimated cost per minute for each transcription model, from $0.003 for gpt-4o-mini-transcribe to $0.006 for whisper-1. The 4o models are metered in audio tokens underneath, so a bill can differ slightly from minutes times the rate.

Does Whisper charge a full minute for a 10-second clip?

OpenAI does not publish a rounding rule. On Modellix, Whisper billed per second: our 5 second clip cost $0.0005 and the 31 second clip $0.0031, at the same $0.0001 per second as a 25 minute file.

What is the Whisper API file size limit?

25 MB per request, in mp3, mp4, mpeg, mpga, m4a, wav, or webm. Split longer recordings into parts under that size; splitting does not add cost.

When does whisper-1 shut down, and what should I use instead?

February 26, 2027, together with the translation endpoint and the gpt-4o transcription models. OpenAI recommends gpt-transcribe ($0.0045 a minute) for recorded audio and gpt-live-transcribe for live audio. Among the models we tested on the same audio, Grok Voice ASR had the lowest list price at $0.10 an hour, Gemini 3.5 Transcribe caught all 20 names and figures, and Fun-ASR and Grok Voice ASR returned speaker turns. None of them returns SRT or VTT directly. All five run on one Modellix API key, so you can test them on your own audio before you switch.

Is gpt-4o-transcribe cheaper than Whisper?

No. Both are $0.006 a minute, and both shut down on February 26, 2027. The cheaper current option from OpenAI is gpt-transcribe at $0.0045 a minute.