OpenAI Whisper API Pricing, and Why I Stopped Paying It (2026)

The Month My “Free” Side Project Cost Me Real Money

In 2025 I had one of those ideas that feels genius at 1 AM and stupid by morning. I’d recorded years of voice memos — rambling product ideas, half-baked feature notes, me arguing with myself in the car about whether to kill Clickly. Hours of it. And I’d shot a bunch of ugly talking-head videos I never published. The idea: transcribe all of it, feed the text into an LLM, and let it spit out blog drafts. A little content machine to feed the distribution grind, which — if you’ve read me before — is the only thing that actually matters after you ship.

The obvious tool was OpenAI’s Whisper. It’s the speech-to-text model everyone reaches for, it’s genuinely excellent, and OpenAI hosts it behind a dead-simple API. Upload audio, get text. I wired it up in an afternoon.

Then I pointed it at the backlog. All of it. In one weekend.

The bill was $43. Not catastrophic. Not “call the bank.” But $43 for a weekend of transcription — for a product making exactly ₹0 — is the kind of thing that makes a broke engineer stare at his Stripe dashboard at 2 AM doing math he should’ve done before wiring up the API. One thought looped in my head: you’re paying per minute for something you could run on your own machine for the price of electricity.

And that’s exactly what Whisper is. Open source. MIT-licensed. Free. OpenAI dumped the weights on GitHub in September 2022. Anyone can download them. The API? It’s just OpenAI renting you their GPUs — a convenience fee that makes perfect sense until the bill lands and your Indian brain converts it to rupees.

Here’s what it actually costs, where it sneaks up on you, and when you should walk away and run it yourself.

---

What OpenAI Charges for Whisper (The Real Numbers)

Unlike a normal LLM that bills you per token, transcription bills you per minute of audio. You pay for how long the recording is, not how much text comes out. Here’s the 2026 lineup, verified against OpenAI’s own pricing:

Model	Price	Per hour of audio
whisper-1 (classic Whisper)	$0.006 / min	~$0.36
gpt-4o-transcribe	~$0.006 / min*	~$0.36
gpt-4o-mini-transcribe	$0.003 / min	~$0.18

*The two newer gpt-4o-*-transcribe models are technically billed by audio tokens ($2.50 / 1M input tokens for the full model), but in practice it works out to roughly the same $0.006/min as classic Whisper for the big one, and about half that for mini. OpenAI rounds to the nearest second, so you’re not paying for a full minute on a 12-second clip.

At a glance, six-tenths of a cent per minute sounds like nothing. And for a while, it is nothing.

Do the math on a light workload: transcribe 100 hours of audio in a month and you’re at $36 on classic Whisper, or $18 on gpt-4o-mini-transcribe. If you’re building a note-taking app where each user records a few voice memos a week, you’ll be swimming in the noise of your other bills. Nobody goes broke here.

The trouble starts when transcription is your product, not a garnish.

---

Where the Per-Minute Bill Actually Bites

Picture the uptime-monitor math I’m always warning about, but for audio. Say you’re building a podcast-summarizing tool, or a meeting-notes SaaS, or — like me — a batch job chewing through a giant archive.

1,000 hours of audio a month is not a wild number for even a small transcription product. A hundred users each uploading a couple of hour-long meetings a week gets you there easily. On classic Whisper that’s $360/month. Per minute it felt free; at volume it’s a car payment, and it scales linearly with every new customer. This is the exact same trap as serverless functions that balloon on you — a number so small you ignore it, multiplied by a volume you didn’t plan for.

Then there are the gotchas that don’t show up on the pricing page:

• The 25 MB file limit. The API rejects anything bigger. An hour of decent-quality audio blows right past it, so you have to chunk long files yourself, transcribe the pieces, and stitch the text back together. That’s engineering time nobody quoted you.
• You pay for silence. Billing is per minute of audio, not per minute of speech. A one-hour recording with forty minutes of dead air still costs you the full hour. Trimming silence before you upload is real money saved.
• Retries cost double. A flaky upload that fails halfway and re-runs? You can get billed for both attempts if you’re not careful about idempotency.
• No free tier worth the name. Unlike email or hosting, there’s no generous “3,000 minutes a month free” cushion to hide under while you find product-market fit.

None of this makes Whisper bad. It’s a fantastic model. It just means the “$0.006, who cares” mindset quietly stops being true right around the time your product starts working.

---

The Part OpenAI Doesn’t Advertise: You Can Run It Yourself

Here’s the reframe that changed everything for me. Whisper isn’t a locked-up API product like GPT-4o. It’s an open-source model you can download and run on your own hardware for the price of electricity. OpenAI’s API is a convenience tax, and a very reasonable one — until it isn’t.

If you’ve read my take on running LLMs locally, you already know this song: the same model that costs money in the cloud runs for free on your laptop if you’re willing to do a little plumbing. Whisper is the friendliest possible version of that story, because the community has already done the hard part.

• faster-whisper — a Python reimplementation by SYSTRAN built on CTranslate2. It’s up to 4x faster than the original OpenAI code for the same accuracy, and uses less memory. This is what I actually use. pip install faster-whisper, point it at a folder, walk away. My entire voice-memo archive — the one that cost me real money on the API — reprocessed overnight on my own machine for nothing.
• whisper.cpp — Georgi Gerganov’s C/C++ port (same person behind the ggml library that powers a huge chunk of the local-AI world). No Python, no CUDA drama, runs beautifully on a Mac’s Apple Silicon. If you want a single fast binary with zero dependency hell, this is it.

You download a model file once — tiny, base, small, medium, or large-v3 depending on how much accuracy you need versus how much RAM you have — and every minute of audio after that is free. For a batch workload, this is a no-brainer. Your electricity bill won’t notice, and there’s no per-minute meter running in the background giving you anxiety.

The catch is self-hosting anything: it’s your machine now. Job crashes at 3 AM? You wake up. Need it running 24/7 as part of a live product? Now you’re babysitting a GPU server, queues, and scaling — and you’ve rebuilt a worse version of the thing you were avoiding. Self-hosting is glorious for batch jobs on your own time. It’s a different animal when it has to be always-on.

---

The Middle Ground: Cheaper Hosted Providers

There’s a third path between “pay OpenAI’s rate” and “run your own server,” and it’s the one most people miss. Whisper being open source means other companies host it too — often far cheaper than OpenAI, because inference speed is their whole business.

• Groq is the wild one. They run whisper-large-v3-turbo on their custom hardware at roughly $0.04 per hour of audio — about a fifteenth of OpenAI’s price — and they do it stupidly fast (hundreds of times faster than real-time). Their bigger whisper-large-v3 is around $0.11/hr. If I were building a transcription product today and didn’t want to babysit a server, Groq is where I’d start. Same open model, tiny fraction of the cost.
• Deepgram runs its own Nova models (not Whisper) at roughly $0.0043/min for batch — cheaper than OpenAI with strong accuracy, and built for high-volume streaming.
• AssemblyAI sits at the premium end, roughly half a cent to just over a cent per minute depending on the tier, but you’re paying for extras — speaker labels, sentiment, chapter detection — baked in.

The point isn’t that any one of these is “the answer.” It’s the same lesson from the Bunny story: if you only know the big brand-name option, you’re limited by your ignorance. Your competitor charging $99/month for transcription might be running the exact same Whisper model you can access for a fraction of the cost — the margin is just knowing the alternatives exist.

---

When to Pay (And When to Walk Away)

I’m not telling you to never pay OpenAI. The hosted API earns its keep for a few very specific situations. If you’re doing a couple hundred minutes a month, paying $0.006/min beats owning a GPU that sits idle — that’s most apps, honestly. If you’re already deep in the OpenAI ecosystem, piping transcripts straight into GPT-4o for summarization, maybe with an agent framework orchestrating the whole thing, one vendor with one bill and one SDK is genuinely worth money. And if you’re a solo founder whose scarcest resource is time, the simplicity of one HTTP call beats everything else on this page. Zero infrastructure, zero model files, zero scaling logic. Just ship.

But flip to self-host or Groq the moment any of this is true: transcription is your product (not a garnish), you’re chewing through big batch jobs on your own schedule, or the monthly bill has crept past what a small GPU or a Groq account costs. That crossover arrives faster than you think. I hit it in one weekend.

For me the answer was embarrassingly simple: faster-whisper on my own machine for the archive, and Groq the day I turned it into a real product. OpenAI’s API got me started and I outgrew it in a weekend — which is exactly the right job for it.

---

Speech-to-text used to be a moat. Now it’s a commodity you can run for free on the laptop you’re reading this on. The Whisper API is a convenience, priced fairly, and genuinely the right call when you’re just testing an idea. But the second it becomes a line item that scales with your success, remember what it actually is underneath: an open-source model OpenAI is renting back to you. Own that, and you own your costs.

I over-engineer everything, I chase every dollar before the product earns one, and every so often that reflex actually pays off. This was one of those times. If you’re piecing together the rest of your stack on a shoestring, the broke solopreneur’s survival guide is the bigger picture.

This is the Broken Engineer Guide. I over-engineer everything, fail at business, and share the scars so you don’t have to earn your own. Go transcribe something.