Search OpenAI’s own pricing page for “whisper-1” today and you will not find it. There is no model by that name on the page, checked 2026-08-22. What is there instead is a family of five transcription models under different names.
Contents
What OpenAI’s pricing page actually shows
Fetched directly from developers.openai.com/api/docs/pricing:
| Model | Use case | Price |
|---|---|---|
| gpt-transcribe | File transcription | $0.0045 / minute |
| gpt-4o-transcribe | File transcription | $0.006 / minute |
| gpt-4o-mini-transcribe | File transcription | $0.003 / minute |
| gpt-realtime-whisper | Live transcription | $0.017 / minute |
| gpt-live-transcribe | Live transcription | $0.017 / minute |
On this specific page, the word “whisper” survives in exactly one model name: the live option gpt-realtime-whisper. Batch file transcription is split across three separate names instead — gpt-transcribe, gpt-4o-transcribe and gpt-4o-mini-transcribe — and the cheapest of the three is gpt-4o-mini-transcribe at $0.003/minute, not gpt-transcribe.
OpenAI’s own changelog logs the change plainly:
> Released GPT Transcribe for accurate file transcription and final transcripts of committed Realtime turns, along with GPT Live Transcribe for low-latency streaming transcription.
Source: developers.openai.com/api/docs/changelog, logged under “Jul 28”.
That entry does not say what happened to whisper-1’s pricing row, and neither does the current pricing page — it simply is not there. So the honest claim is narrower than “Whisper got replaced”: the current pricing page does not price whisper-1, and OpenAI now lists three differently-named, differently-priced models for the same batch-transcription job instead.
The other “Whisper” — the one you can run yourself
This is a separate thing from the API, and worth keeping separate. The original Whisper model is open-weight, hosted on GitHub, and not part of OpenAI’s hosted-API pricing at all:
> Whisper’s code and model weights are released under the MIT License.
Source: github.com/openai/whisper
Nothing in the pricing-page change above touches this license. The open-weight model still exists, and nothing stops you from running it on your own infrastructure — the MIT license doesn’t remove the cost of doing that, it just moves the cost from a per-minute API fee to your own compute and upkeep.
I checked Deepgram’s own model documentation for an equivalent statement about Nova-3 and did not find one. It describes the model as reachable through Deepgram’s API endpoint and says nothing about the weights being open, self-hostable, or proprietary. I could not find a self-hosted Nova-3 deployment page either. That is not the same as “Nova-3 is closed” — it means I could not confirm a licensing position for it the way I could for Whisper, so this piece does not claim one.
Deepgram Nova-3’s numbers
Fetched directly from deepgram.com/pricing:
| Model | Mode | Price |
|---|---|---|
| Nova-3 Monolingual | Pre-recorded, pay-as-you-go | $0.0043 / minute |
| Nova-3 Monolingual | Streaming, current promotional rate | $0.0048 / minute |
| Nova-3 Monolingual | Streaming, listed regular rate | $0.0077 / minute |
| Whisper Large (Deepgram-hosted) | Both plans shown | $0.0048 / minute |
Two things worth noticing in that table. First, Deepgram’s own streaming price is explicitly split into a current promotional rate and a separate regular rate — the page states this directly:
> No minimums. No expiration. No credit card required.
Source: deepgram.com/pricing
That line describes the pay-as-you-go plan structure, not a permanent price. If you build a cost model off the promotional $0.0048/min streaming rate, check the page again before it ships — “current” is doing real work in that sentence.
Second: Deepgram hosts Whisper Large itself, priced at $0.0048/min in both plan columns on its pricing page. So the open-weight Whisper model isn’t only available through OpenAI or your own hardware — you can also rent it through Deepgram’s infrastructure, at the same listed price Deepgram charges for its own Nova-3 Monolingual streaming promotional rate.
What this comparison does not tell you
It does not tell you which one transcribes more accurately. I looked for a word-error-rate figure on OpenAI’s own model page and on Deepgram’s own pricing and model pages, and neither publishes a number I could quote and verify on the spot. Third-party benchmark posts claim double-digit error-rate improvements for gpt-transcribe over whisper-1, but I could not trace that number back to a primary source I could check, so it is not in this piece. If you need an accuracy comparison, run both APIs on your own audio — the ten minutes that takes is worth more than a WER number from a page that changes the models under it every few months.
Picking between them
Self-hosting need decides the first branch — that is a licensing and infrastructure question, not a pricing one, and Whisper’s MIT license is the only confirmed open option on that side. Past that, it comes down to whether the job is live audio or a finished file, and the batch and live prices don’t move together.
For a finished file, the pay-as-you-go rates from lowest to highest are: gpt-4o-mini-transcribe at $0.003/min, Nova-3 Monolingual pre-recorded at $0.0043/min, gpt-transcribe at $0.0045/min, Deepgram-hosted Whisper Large at $0.0048/min, then gpt-4o-transcribe at $0.006/min. None of these include volume discounts — Deepgram’s Growth tier brings Nova-3 Monolingual pre-recorded down to $0.0036/min, but that rate requires pre-paid annual credits, not a pay-as-you-go account, so it isn’t a like-for-like comparison with the numbers above it.
For live audio, the two vendors are not close. OpenAI’s live options — gpt-realtime-whisper and gpt-live-transcribe — are both $0.017/min. Deepgram’s Nova-3 streaming is $0.0048/min at its current promotional rate, or $0.0077/min at the regular rate the page states will apply once the promotion ends. Even at the regular rate, OpenAI’s live transcription lists at more than twice Deepgram’s.
If you are budgeting a project that mixes transcription with LLM calls — summarizing the transcript, for example — the per-minute audio cost is only half the bill. A token counter for estimating OpenAI and Claude API costs covers the other half without guessing at it.
This article was written with the help of AI. Every price and quoted sentence above was pulled from OpenAI’s and Deepgram’s own pages on 2026-08-22 and checked against those pages before publication. The first draft named gpt-transcribe as the cheapest OpenAI batch option and called the two live-transcription rates close to each other — both wrong once checked against the same numbers already in this piece’s own tables, and corrected here.