The short answer
Short answer, prices read on 21 October 2026: a transcribed hour of recorded audio costs $0.15 on AssemblyAI Universal-2, $0.27 on OpenAI‘s gpt-transcribe and $0.258 on Deepgram Nova-3 pay as you go. A 5-person sales team that records 100 hours a month pays $15 to $27 a month on those APIs, against $84.95 for five Otter.ai Pro seats billed monthly. Self-hosting Whisper with faster-whisper on a rented GPU costs well under a cent of GPU time per audio hour when the GPU runs only for the batch, but a GPU left on all month ($194.40) only beats gpt-transcribe above 720 audio hours a month. Seat tools cost more per hour because they sell notes, summaries and meeting bots, not just transcripts.
| Option | Price as listed | Per audio hour | 100 hours a month |
|---|---|---|---|
| AssemblyAI Universal-2 | $0.15 an hour | $0.15 | $15 |
| OpenAI gpt-4o-mini-transcribe (removed 26 February 2027) | $0.003 a minute | $0.18 | $18 |
| AssemblyAI Universal-3.5 Pro | $0.21 an hour | $0.21 | $21 |
| Deepgram Nova-3 Monolingual | $0.0043 a minute pay as you go, $0.0036 on Growth | $0.258 / $0.216 | $25.80 |
| OpenAI gpt-transcribe | $0.0045 a minute | $0.27 | $27 |
| Deepgram hosted Whisper Large | $0.0048 a minute | $0.288 | $28.80 |
| OpenAI whisper-1 or gpt-4o-transcribe (removed 26 February 2027) | $0.006 a minute | $0.36 | $36 |
| HyNote Pro, billed yearly Our pick | $79.99 a year, 1,200 minutes a month | $0.3333 at the full allowance | $33.33 (5 seats) |
| Otter.ai Pro, billed annually | $8.33 a user a month, 1,200 minutes | $0.4165 at the full allowance | $41.65 (5 seats) |
| Otter.ai Pro, billed monthly | $16.99 a user a month, 1,200 minutes | $0.8495 at the full allowance | $84.95 (5 seats) |
| Whisper via faster-whisper on a RunPod RTX A5000, GPU on only for the batch | $0.27 a GPU hour | about $0.0059 of GPU time | about $0.59 of GPU time |
Our pick means we earn a commission if you subscribe through our link. HyNote is our pick for teams that want meeting notes and summaries, not raw transcripts, because its Pro plan is $0.3333 an hour at the full allowance against $0.4165 for Otter Pro billed annually. That does not change any score: HyNote scores 6.0 on our published rubric and Otter.ai 6.2. See HyNote plans.
Is Whisper transcription free?
The software is. OpenAI’s README says “Whisper’s code and model weights are released under the MIT License”, and the two projects most people run it with in production, faster-whisper and whisper.cpp, are MIT licensed too. You can use all three commercially at no licence cost. What you pay for is the computer that runs the model, or a vendor that runs it for you.
| Project | Licence | GitHub stars | Latest release | Last push |
|---|---|---|---|---|
| openai/whisper | MIT | 110,977 | v20250625 (26 June 2025) | 31 August 2026 |
| SYSTRAN/faster-whisper | MIT | 26,009 | v1.2.1 (31 October 2025) | 16 October 2026 |
| ggml-org/whisper.cpp | MIT | 54,531 | v1.9.5 (6 October 2026) | 19 October 2026 |
OpenAI sells no Whisper subscription. The sites that rank for “whisper subscription” are third-party apps built on the model. For a desktop note taker that transcribes locally with no per-minute bill, Meetily is an open-source option in our catalogue.
What OpenAI charges for transcription, and the February 2027 cutoff
OpenAI’s pricing page lists five transcription models. Three are for recorded files; two are for live audio and cost more than three times as much per minute.
| Model | Use | Per minute | Per audio hour | Status |
|---|---|---|---|---|
| gpt-transcribe | Files | $0.0045 | $0.27 | Recommended replacement |
| gpt-4o-transcribe | Files | $0.006 | $0.36 | Removed 26 February 2027 |
| gpt-4o-mini-transcribe | Files | $0.003 | $0.18 | Removed 26 February 2027 |
| whisper-1 | Files | $0.006 | $0.36 | Removed 26 February 2027; priced on its model page, not in the pricing table |
| gpt-live-transcribe | Live | $0.017 | $1.02 | Recommended replacement |
| gpt-realtime-whisper | Live | $0.017 | $1.02 | Listed |
On 26 August 2026 OpenAI notified developers that whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe and gpt-4o-transcribe-diarize will be removed from the API on 26 February 2027, and names gpt-live-transcribe or gpt-transcribe as the replacement for each. The budget effect is real: the $0.18 an hour model goes away, so OpenAI’s floor for files becomes gpt-transcribe at $0.27. On 100 hours a month that is $9 more. Anyone who built on whisper-1 at $0.36 an hour saves $0.09 an hour by moving to gpt-transcribe.
What self-hosting costs: GPU time, speed, and the idle trap
The faster-whisper README publishes one benchmark: 13 minutes of audio, the Large-v2 model, fp16, beam size 5, on an NVIDIA RTX 3070 Ti 8GB. openai/whisper took 2m23s, whisper.cpp 1m05s, faster-whisper 1m03s, and faster-whisper with batch_size=8 took 17 seconds. That is 45.9 times real time batched and 12.4 times unbatched, or 78.5 and 290.8 GPU seconds per audio hour. We priced those GPU seconds at RunPod‘s listed Pod rates.
| Setup | GPU rate | GPU time per audio hour | Cost per audio hour |
|---|---|---|---|
| faster-whisper, batch 8, RTX A5000 | $0.27 an hour | 78.5 s | $0.0059 |
| faster-whisper, batch 8, RTX 4090 | $0.89 an hour | 78.5 s | $0.0194 |
| faster-whisper, unbatched, RTX A5000 | $0.27 an hour | 290.8 s | $0.0218 |
| faster-whisper, unbatched, A40 | $0.59 an hour | 290.8 s | $0.0477 |
| Any of the above, pod left on all month | $0.27 x 24 x 30 | n/a | $194.40 a month, flat |
Three things the per-hour figure leaves out. First, it counts GPU time only: no engineer time to build the queue, no storage, no egress, no failed jobs. Second, the speed was measured on a 3070 Ti, not on the GPUs priced here. Third, and most important, idle time. A pod that stays on all month costs $194.40 whether it transcribes one hour or a thousand. At that flat cost, self-hosting only matches gpt-transcribe at 720 audio hours a month ($194.40 / $0.27) and AssemblyAI Universal-2 at 1,296 hours ($194.40 / $0.15). Below that, start the GPU for each batch and stop it afterwards, or use an API. The same break-even logic applies to language models; see Self-hosting an LLM against paying per token: the break-even.
What a team actually pays, by volume
| Who | Hours a month | gpt-transcribe | AssemblyAI Universal-2 | Deepgram Nova-3 | Self-host GPU time |
|---|---|---|---|---|---|
| Weekly podcast | 4 | $1.08 | $0.60 | $1.03 | under $0.10 |
| 5-person sales team | 100 | $27 | $15 | $25.80 | $0.59 |
| Agency transcribing client calls | 200 | $54 | $30 | $51.60 | $1.18 |
Two vendor terms move these totals. AssemblyAI bills multichannel audio per channel, so a 1-hour stereo recording is billed as 2 hours, which makes Universal-2 $0.30 an hour for a dual-channel sales call. And its free tier includes up to 185 hours of pre-recorded transcription: at 4 hours a month that is 46 months of a podcast, though the page does not say whether the allowance renews.
What switching off seat tools saves, and what you give up
Five Otter.ai Pro seats on monthly billing cost $84.95 a month, or $1,019.40 a year, for up to 100 hours of transcription. The same 100 hours cost $15 on AssemblyAI Universal-2, so the spend avoided is $69.95 a month ($839.40 a year). On gpt-transcribe it is $57.95 a month, and on a self-hosted GPU run only for the batch it is about $84.36 a month before engineer time. On Otter’s annual billing the five seats are $41.65, and on HyNote Pro yearly they are $33.33, so the gap to an API narrows to a few dollars.
What the API price does not include: a meeting bot that joins the call, speaker-labelled notes, summaries, action items and a searchable library for the team. With an API, someone has to upload the recordings and turn the transcript into notes. If the team wants notes, a seat tool is the product; if it wants text to feed another system, an API or self-hosted Whisper is. For the note-taker side of this choice, see AI note takers compared, with scores and verified prices and Otter.ai alternatives and what each one costs.
Where search results get it wrong
We captured Google’s AI Overview and AI Mode for “whisper transcription cost” on 19 October 2026 and checked each claim against OpenAI’s pages, re-read on 21 October 2026.
| Search result says | Vendor page says |
|---|---|
| AI Overview: OpenAI deprecated whisper-1 on 26 August 2026 | 26 August 2026 is the date OpenAI notified developers. Removal is 26 February 2027, and the same notice covers gpt-4o-transcribe, gpt-4o-mini-transcribe and gpt-4o-transcribe-diarize |
| AI Mode: gpt-transcribe costs $0.017 a minute | $0.017 a minute is gpt-live-transcribe and gpt-realtime-whisper, the live models. gpt-transcribe is $0.0045 a minute |
| AI Overview and AI Mode: whisper-1 at $0.006 a minute is OpenAI’s price | whisper-1 is still $0.006 on its model page, but it is not in the pricing page’s transcription table and is removed on 26 February 2027 |
Which one to pick
| If you | Pick | Why |
|---|---|---|
| Want meeting notes and summaries for a few people | HyNote Pro Our pick | $0.3333 an hour at the full 1,200 minutes, notes included. See HyNote plans |
| Already live in Otter | Otter.ai Pro, annual billing | $0.4165 an hour against $0.8495 on monthly billing |
| Need transcripts for another system, under about 700 hours a month | AssemblyAI Universal-2 or OpenAI gpt-transcribe | $0.15 to $0.27 an hour, no servers |
| Have an engineer and batch jobs | faster-whisper on a rented GPU, started per batch | Cents per audio hour in GPU time; MIT licence |
| Run a GPU all month anyway | Self-host on that GPU | Flat $194.40 a month beats gpt-transcribe above 720 hours |
The tools on this page
HyNote Our pick
AI note taker for meetings, voice and PDFs. Pro is $79.99 a year ($6.66 a month shown) with 1,200 transcription minutes a month; Teams is $15 a seat a month on yearly billing; the free plan records up to 120 minutes a session; paid plans have a 7-day free trial. Our full test is in HyNote review: what 1,200 minutes a month costs. Try HyNote.
Otter.ai
Meeting agent with real-time transcription and summaries. Pro is $16.99 a user a month billed monthly or $8.33 billed annually, with 1,200 minutes a user a month and no rollover; Basic is free with 300 minutes; Business is $24 a user a month billed monthly.
faster-whisper
MIT-licensed reimplementation of Whisper on CTranslate2. Its README benchmark runs 13 minutes of audio in 17 seconds batched on an RTX 3070 Ti, against 2m23s for openai/whisper on the same GPU.
Meetily
Open-source meeting note taker that transcribes on your own machine, so there is no per-minute bill. It is the self-hosted answer for teams that want notes rather than raw transcripts.
OpenAI, AssemblyAI, Deepgram, RunPod and whisper.cpp have no tool page on this site; they link to their own pricing pages or repositories. Granola is another bot-free note taker in the same category.
How we checked
Every price on this page was read on the vendor’s own page in a rendered browser on 21 October 2026: OpenAI’s pricing, whisper-1 and deprecations pages, AssemblyAI, Deepgram (pre-recorded tab, both plans), HyNote, Otter (monthly and annual), and RunPod. Licences, stars and releases come from each repository’s README and the GitHub REST API the same day. The benchmark is the faster-whisper README’s own table. Team sizes, hours and every total are our arithmetic. We did not run our own transcription benchmark or compare accuracy between models, so this is a price check, not an accuracy test. For more self-hosted replacements for paid tools, see Open source alternatives to paid AI tools, and the monthly cost saved.
Questions people ask about Whisper transcription cost
How much does a Whisper transcription cost per hour?
On OpenAI’s API, whisper-1 and gpt-4o-transcribe cost $0.006 a minute, or $0.36 an hour; gpt-transcribe is $0.27 an hour and gpt-4o-mini-transcribe $0.18. Deepgram’s hosted Whisper Large is $0.288 an hour. Prices read 21 October 2026.
Is whisper-1 being shut down?
Yes. OpenAI notified developers on 26 August 2026 that whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe and gpt-4o-transcribe-diarize will be removed from the API on 26 February 2027. The recommended replacement is gpt-live-transcribe or gpt-transcribe.
Is Whisper transcription free?
The code and model weights are MIT licensed, so the software is free, including for commercial use. You pay for the computer that runs it, or for an API that runs it for you.
How much is a Whisper subscription?
OpenAI sells no Whisper subscription. Its API is pay as you go per minute. Apps sold as Whisper subscriptions are third-party products built on the open model.
What does it cost to run Whisper on my own GPU?
Using the faster-whisper README benchmark and RunPod’s $0.27 an hour RTX A5000 rate, batched transcription is about $0.0059 of GPU time per audio hour if the GPU runs only for the job. A pod left on all month costs $194.40, which only matches gpt-transcribe at 720 audio hours a month.
What is the cheapest transcription service?
Among the APIs we priced on 21 October 2026: AssemblyAI Universal-2 at $0.15 an hour, then OpenAI gpt-4o-mini-transcribe at $0.18 until its removal on 26 February 2027, Deepgram Nova-3 at $0.216 on Growth and OpenAI gpt-transcribe at $0.27. Self-hosted Whisper costs less in GPU time but needs an engineer.
Do note takers like Otter or HyNote replace a transcription API?
They do a different job: meeting bots, notes and summaries on top of the transcript. At the full 1,200 minutes a month, HyNote Pro yearly works out to $0.3333 an hour and Otter Pro $0.4165 annual or $0.8495 monthly, against $0.15 to $0.27 for a raw transcript from an API.
Related reading
- HyNote review: what 1,200 minutes a month costs
- AI note takers compared, with scores and verified prices
- Otter.ai alternatives and what each one costs
- Self-hosting an LLM against paying per token: the break-even
- Open source alternatives to paid AI tools, and the monthly cost saved
Sources
- OpenAI API pricing (transcription models), read 21 October 2026.
- OpenAI, whisper-1 model page, read 21 October 2026.
- OpenAI API deprecations (2026-08-26 transcription models), read 21 October 2026.
- AssemblyAI pricing, read 21 October 2026.
- Deepgram pricing (pre-recorded), read 21 October 2026.
- HyNote pricing, read 21 October 2026.
- Otter.ai pricing (monthly and annual), read 21 October 2026.
- RunPod pricing (Pods), read 21 October 2026.
- openai/whisper on GitHub (README, REST API), read 21 October 2026.
- SYSTRAN/faster-whisper on GitHub (README benchmark, REST API), read 21 October 2026.
- ggml-org/whisper.cpp on GitHub (REST API), read 21 October 2026.
- Google AI Overview and AI Mode for “whisper transcription cost”, our capture, 19 October 2026.
Comments
Sign in to join the discussion.
Login to comment