TL;DR: Cloud transcription is cheap in a way cloud LLMs are not — $0.15 to $0.36 per audio hour as of October 2026. A used RTX 3060 running faster-whisper costs $260–$296 up front and about half a cent per audio hour after that, so it needs roughly 780–1,870 transcribed hours to pay for itself. Self-host for privacy or bulk volume, not to save $5 a month.

	Self-hosted (faster-whisper + used GPU)	Cloud batch API (AssemblyAI / Deepgram)	OpenAI whisper-1 / gpt-4o-transcribe
Best for	Audio that can’t leave the building; 300+ hrs/mo; big backlogs	Low volume, zero ops, built-in diarization	Teams already on the OpenAI stack
Cost	$260–$296 once + ~$0.006/audio-hr electricity	$0.15–$0.26 per audio hour	$0.36 per audio hour
The catch	780–1,870 hours to break even; you run the pipeline	Your audio transits a third party	Priciest mainstream rate; 25 MB per-file limit

Honest take: under ~100 hours a month with nothing sensitive in the audio, use AssemblyAI or gpt-4o-mini-transcribe and keep your $280. The moment recordings contain patient calls, legal interviews, or anything under NDA — or your volume passes a few hundred hours a month — a used RTX 3060 is the better deal and pays off in months.

Every price below was checked in October 2026. Speech-to-text pricing moved several times in 2026 (Deepgram’s Nova-3 rates in particular), so re-verify before budgeting a year out.

What does cloud transcription actually cost in 2026?

The three APIs most self-hosters benchmark against, at their published pay-as-you-go rates for pre-recorded audio, as of October 2026:

Provider / model	Per minute	Per audio hour	Notes
AssemblyAI Universal-2 (batch)	$0.0025	$0.15	$50 free credit; billed per second
OpenAI gpt-4o-mini-transcribe	$0.003	$0.18	Token-billed; ~$0.003/min per OpenAI’s own estimate
Deepgram Nova-3 (batch, monolingual)	$0.0043	$0.26	$200 free credit; streaming is $0.0077/min regular
OpenAI whisper-1 / gpt-4o-transcribe	$0.006	$0.36	25 MB file limit forces chunking on long recordings

Two caveats on that table. Deepgram’s pricing was in flux through late 2026 — some trackers list a promotional $0.0048/min streaming rate against a $0.0077 regular price, and the older $0.0043 batch rate was being phased toward Nova-3’s newer tiers, so treat the Deepgram row as a floor. And add-ons stack: speaker diarization adds roughly $0.002/min at Deepgram, and AssemblyAI’s extras (PII redaction, entity detection) each carry their own per-hour fee. A “cheap” API call with three add-ons can double.

What those rates mean monthly:

Monthly audio	AssemblyAI ($0.15/hr)	Deepgram batch ($0.26/hr)	whisper-1 ($0.36/hr)	Self-hosted RTX 3060 (power only)
10 hrs	$1.50	$2.60	$3.60	~$0.06
100 hrs	$15	$26	$36	~$0.60
500 hrs	$75	$129	$180	~$2.90
2,000 hrs	$300	$516	$720	~$11.60

The left three columns are why this decision differs from the LLM one. Self-hosting a chat model replaces a $20–$200/month subscription. Self-hosting transcription, at typical personal volumes, replaces a bill smaller than a coffee.

What does self-hosted Whisper cost to run?

The software side is genuinely free: OpenAI’s Whisper weights, faster-whisper (the CTranslate2 reimplementation most people should use), and whisper.cpp are all MIT-licensed with no commercial restrictions.

The hardware floor is low. Whisper large-v3 in int8 runs in under 4 GB of VRAM in most community reports (fp16 figures range from ~3 to ~6 GB depending on backend and batch settings), so any 8 GB card works with headroom. The value pick in October 2026 is a used RTX 3060 12GB at $260–$296 — the 12 GB buffer also lets the same card serve a small LLM when it isn’t transcribing. A new RTX 4060 8GB runs $297–$330 and transcribes slightly faster.

Running cost is electricity and nothing else. At roughly 5× real-time (conservative for a 3060-class card, see the speed section below), 100 hours of audio is 20 GPU-hours. At ~170 W draw and the US-average ~$0.17/kWh, that’s about $0.60 a month — half a cent per audio hour. Regional rates move that between $0.35 and $1.40; our electricity cost by region breakdown has the full table. Unlike an always-on LLM server, a transcription box only draws GPU power while a job is running.

Setup is a pip install:

pip install faster-whisper

from faster_whisper import WhisperModel

model = WhisperModel("large-v3", device="cuda", compute_type="int8_float16")
segments, info = model.transcribe("meeting.mp3")
print(f"Detected {info.language} (p={info.language_probability:.2f})")
for s in segments:
    print(f"[{s.start:6.1f}s -> {s.end:6.1f}s] {s.text}")

Expected output looks like:

Detected en (p=0.99)
[   0.0s ->    4.2s]  Okay, let's get started with the quarterly review.
[   4.2s ->    9.8s]  First item is the migration timeline we pushed last sprint.

If you’re choosing between the three local runtimes, our faster-whisper vs whisper.cpp vs WhisperX comparison covers the trade-offs; the short version is faster-whisper for GPU batch work, WhisperX when you need word-level timestamps and diarization, whisper.cpp for CPU-only machines.

Where is the break-even?

Divide the card price by the API rate you’d otherwise pay. Using the $280 midpoint for a used RTX 3060 and ignoring electricity (it shifts these numbers by under 2%):

API you’d be replacing	Rate	Audio hours to pay off the card
AssemblyAI Universal-2	$0.15/hr	~1,870 hrs
gpt-4o-mini-transcribe	$0.18/hr	~1,560 hrs
Deepgram Nova-3 batch	$0.26/hr	~1,080 hrs
OpenAI whisper-1	$0.36/hr	~780 hrs

Translate that to time at your volume:

• 20 hrs/month (meeting notes, a weekly podcast): 4–8 years to break even against the cheap APIs. Don’t buy hardware for this.
• 150 hrs/month (a small podcast network, research interviews): 7–12 months against the cheap APIs, ~5 months against whisper-1. Defensible, not compelling — unless privacy is in play, which flips it immediately.
• 500+ hrs/month (call recordings, media monitoring, meeting capture for a team): 2–4 months. Self-hosting wins on cost alone, and the card keeps earning after payback.

One scenario breaks the formula: the one-time backlog. If you have 1,000 hours of archive audio and no recurring volume, don’t buy a card and don’t pay $150–$360 in API fees. Rent. A 3090-class GPU on Vast.ai starts around $0.07/hr (market price, October 2026 — it floats). Even at a conservative 8× real-time, 1,000 audio hours is ~125 GPU-hours: roughly $9–$15 of rented compute for a job that costs $360 at whisper-1 rates. That’s a 25–40× saving with zero hardware commitment.

How fast is local Whisper on consumer GPUs?

Fast enough that throughput is never the bottleneck at home-lab scale. Community benchmarks for faster-whisper large-v3 in int8 cluster around 4–6× real-time on an RTX 4060 and 12–16× real-time on an RTX 4090 (two independent 2026 benchmark sets agree within rounding; exact numbers move with beam size and batch settings). A used RTX 3060 lands near the 4060; a 3090 sits well above it.

At 5× real-time, one mid-range card clears 2,000 audio hours a month running 13 hours a day. There is no realistic personal or small-team workload that outgrows a single GPU.

CPU-only is the real constraint to respect: large-v3 through whisper.cpp runs slower than real-time on most desktop CPUs, so the small and medium models are the practical ceiling without a GPU. That works for note-taking; it does not work for a 500-hour backlog.

What do the cloud APIs buy you that local Whisper doesn’t?

Paying $0.15–$0.36/hr gets you real things, and pretending otherwise is how self-hosting advice loses credibility:

• Diarization and formatting out of the box. Deepgram and AssemblyAI return speaker labels, punctuation, and timestamps from one API call. Locally that means wiring WhisperX plus pyannote, which is an afternoon of setup and another model to maintain.
• Streaming at scale. Live captioning with sub-second latency across many concurrent streams is genuine infrastructure. Local streaming Whisper works for one or two streams; beyond that you’re building a product, not running a tool.
• Accuracy on hostile audio. On clean recordings, large-v3 is competitive with anything commercial. On 8 kHz telephony and heavy crosstalk, Deepgram’s and AssemblyAI’s purpose-trained models have the edge in most published comparisons — if call-center audio is your workload, benchmark both on your own recordings before deciding anything.
• Nobody on call. Whisper large-v3 has a documented habit of hallucinating text on long silences (openai/whisper issue tracker; mitigated by VAD filtering, which faster-whisper ships built-in but you have to enable). With an API, that’s the vendor’s problem.

When NOT to self-host transcription

• Your volume is under ~50 hours a month and the audio isn’t sensitive. The math above is unambiguous: break-even sits years out, and $50–$200 of free API credit covers months of light use before you pay anything.
• You need live captions for events or multi-stream real-time. Streaming rates are higher ($0.45/hr at AssemblyAI, $0.46/hr regular at Deepgram), but the infrastructure you’d replace is much more than a GPU and a script.
• You need diarization, redaction, and compliance features yesterday. Assembling the local equivalent is a real project. Ship with the API; revisit when volume justifies the build.
• You have no GPU and no appetite for one. CPU-only large-v3 is painful, and buying a card purely for sub-100-hour monthly volume never pays back. This is the same discipline we apply to LLMs in when NOT to self-host AI: self-hosting is a tool, not an identity.

The privacy exception overrides all four: if recordings legally or contractually cannot transit a third party — medical, legal, anything under NDA — the cloud rows above are not options at any price, and the decision collapses to which local setup to run.

What to actually buy

Prices as of October 2026, all verified in the comparison above:

Your situation	Pick	Price	Where
Under ~50 hrs/mo, nothing sensitive	AssemblyAI or gpt-4o-mini-transcribe API	$0.15–$0.18/audio-hr	Their pricing pages; both have free credit
One-time backlog, no recurring volume	Rented 3090-class GPU + faster-whisper	from $0.07/GPU-hr	Vast.ai
Ongoing 150+ hrs/mo, or any privacy requirement	Used RTX 3060 12GB	$260–$296	Check price
Same, but new card with LLM headroom too	RTX 4060 Ti 16GB	$450–$853	Check price
Already own an 8GB+ GPU	faster-whisper on it, int8	$0	—

FAQ

Is Whisper free for commercial use? Yes. The Whisper model weights, faster-whisper, and whisper.cpp are all MIT-licensed as of October 2026 — no usage caps, no commercial carve-outs. Transcripts you generate are yours.

How much VRAM does Whisper large-v3 need? Under 4 GB in int8 by most community reports, and roughly 3–6 GB in fp16 depending on backend and batch settings. Any 8 GB consumer GPU runs it comfortably; a 12 GB card adds room to batch files or co-host a small LLM.

Is cloud transcription cheaper than self-hosting? At low volume, yes — decisively. A $280 used GPU needs roughly 780–1,870 transcribed hours to match what you’d have spent on APIs at October 2026 rates ($0.15–$0.36/audio-hr). Past ~150 hours a month, or with any requirement that audio stays on-premises, self-hosting wins.

Sources

• OpenAI API pricing — whisper-1 and gpt-4o-transcribe rates, accessed October 2026
• Deepgram pricing — Nova-3 pay-as-you-go rates, accessed October 2026
• AssemblyAI pricing — Universal model rates, accessed October 2026
• faster-whisper (GitHub) — CTranslate2 Whisper implementation, MIT license
• openai/whisper (GitHub) — model weights and license

Recommended Gear

• RTX 3060 12GB — $260–$296 used; the break-even pick for ongoing transcription volume
• RTX 4060 Ti 16GB — $450–$853 new; transcription plus room for a 13B-class LLM

What self-hosting actually costs

Real cost breakdowns for self-hosted AI: hardware floors, power, maintenance hours, and the honest comparison against paying for it. No spam, unsubscribe anytime.

Email address Subscribe