You need speech-to-text in your app. Whisper Large V3 keeps showing up as the answer - 99 languages, solid accuracy, MIT license. The model itself is settled science. What isn't settled is where you run it.
OpenAI hosts it at $0.36/hour. Groq runs a turbo variant for $0.02/hour. Deepgram built their own model that beats Whisper on noisy audio. AssemblyAI bundles diarization and sentiment analysis on top. deAPI transcribes directly from YouTube URLs for $0.021/hour. And you can always self-host the thing on your own GPU.
This article compares all six options on the metrics that actually drive the decision: price per hour of audio, speed, features you get out of the box, and the integration quirks nobody mentions until you're knee-deep in code.
The pricing table you came here for
Every price below is list rate as of June 2026. Enterprise discounts, volume tiers, and committed-use agreements can drop these 30-70% - but most developers reading this aren't negotiating enterprise contracts.
| Provider | Model | Price/hour | Billing model |
|---|---|---|---|
| OpenAI | Whisper large-v3 | $0.36 | Per minute ($0.006/min) |
| Groq | Whisper large-v3-turbo | ~$0.02 | Per hour |
| Deepgram | Nova-3 | $0.26 (batch) / $0.46 (stream) | Per minute |
| AssemblyAI | Universal-2 | $0.12 (Nano) / $0.75 (Best) | Per minute |
| deAPI | Whisper large-v3 | $0.021 | Per hour of audio |
| Self-hosted | Whisper large-v3 | $0.05-0.15 (GPU cost) | Your infrastructure |
The spread is 17x between the cheapest hosted option and the most expensive. Same underlying model architecture, radically different price tags. The difference comes from hardware (consumer GPUs vs. cloud A100s), billing granularity, and what's bundled in.
What each option actually gives you
OpenAI Whisper API
Most developers start here. Upload a file, get a transcript - the SDK and docs have been battle-tested for years, and Stack Overflow covers every edge case.
The simplicity has a ceiling, though. Streaming and speaker diarization don't exist. The 25 MB file size cap forces you to chunk long recordings, then stitch transcripts back together on your side. Processing speed sits around 45-60 seconds per hour of audio.
At $0.36/hour, OpenAI charges 17x more than the cheapest hosted alternative. That gap is invisible when you're transcribing a few test files. Cross 100 hours per month and it's $36 that could be $2.10 on deAPI.
The sweet spot: quick integration, prototyping, and teams already deep in the OpenAI ecosystem who value familiarity over cost.
Groq Whisper
Groq runs Whisper large-v3-turbo on custom LPU hardware. One hour of audio transcribes in 8-12 seconds. Price matches the speed: ~$0.02/hour.
You give up the same things as with OpenAI (streaming, diarization, 25 MB file cap), plus Groq adds its own wrinkle: availability drops during peak demand, and the free tier rate limits are tight enough to block serious testing.
Where it shines: batch pipelines that need to chew through hundreds of hours overnight. Podcast archives, meeting backlogs, content indexing - anything where latency to the end user doesn't matter.
Deepgram Nova-3
Deepgram didn't just host Whisper - they built Nova-3 from scratch. On clean English, it matches Whisper. On noisy, accented, and phone-quality audio, it pulls ahead: ~9.4% WER on telephony vs. Whisper's ~12.8%.
Batch transcription costs $0.26/hour. Streaming runs $0.46/hour but delivers sub-300ms latency with real-time diarization. The $200 free credit on signup covers a full evaluation.
AssemblyAI
AssemblyAI sells the layer above transcription. Universal-2 handles 99 languages with diarization, and "Audio Intelligence" add-ons let you bolt on sentiment analysis, PII redaction, topic detection, and summarization per job.
Read the pricing carefully, though. Nano ($0.12/hour) covers basic transcription. Best ($0.75/hour) improves accuracy. Each add-on stacks $0.02-0.08/hour extra, so a fully-featured pipeline can double the headline rate before you notice.
The $50 credit plus 185 free hours gives you real runway for testing. Meeting assistants, compliance workflows, content analysis platforms - anything where raw text isn't enough and you need structured intelligence on top.
deAPI
deAPI runs Whisper Large V3 on a distributed network of consumer-grade GPUs. The price reflects that architecture: $0.021 per hour of audio, which makes it the cheapest hosted Whisper endpoint that runs the full (non-turbo) model.
The standout feature is direct URL transcription. Pass a YouTube, Twitch, TikTok, Kick, or X URL - including X Spaces - and the API handles audio extraction server-side. You skip the yt-dlp → ffmpeg → format conversion → chunking pipeline entirely, which saves more engineering time than the pricing difference suggests.
import requests
response = requests.post(
"https://api.deapi.ai/api/v1/client/transcribe",
headers={"Authorization": "Bearer YOUR_KEY"},
data={
"source_url": "https://www.youtube.com/watch?v=VIDEO_ID",
"model": "WhisperLargeV3",
"include_ts": True
}
)
request_id = response.json()["data"]["request_id"]
Six lines of Python. The URL goes in, the transcript comes back with timestamps. Compare that to the typical Whisper pipeline: download video with yt-dlp, extract audio with ffmpeg, convert to the right format, chunk if over 25 MB, upload, transcribe, stitch chunks back together. deAPI collapses that entire chain into a single API call.
The API also supports OpenAI SDK compatibility, so migrating from OpenAI's Whisper endpoint means changing base_url and api_key while keeping your existing parsing logic intact.
The initial_prompt parameter works the same way as OpenAI's - a text snippet that conditions the model toward specific terminology, proper nouns, and formatting conventions.
Limitations: Batch only - no streaming, no diarization. A 10-minute video typically processes in under 30 seconds.
The migration path from OpenAI is the easiest of any provider here: swap two lines of code, keep your parsing logic, cut your bill by 17x.
Self-hosted (faster-whisper / whisper.cpp)
Running Whisper on your own GPU eliminates per-minute costs entirely. faster-whisper delivers 4-10x speedups over the original implementation; whisper.cpp runs on CPU if you're patient.
A cloud L4 instance costs $0.05-0.15/hour depending on provider. At high volume, transcription cost per hour approaches zero because you're paying for the GPU regardless of utilization.
The bill you don't see is engineering time. GPU provisioning, 25 MB chunking logic, hallucination mitigation on silent segments, deployment maintenance - each one is a small project that never fully goes away. Diarization means bolting on pyannote as a separate pipeline.
Makes sense at 1000+ hours/month, or in air-gapped environments where API calls aren't an option.
Feature comparison
| Feature | OpenAI | Groq | Deepgram | AssemblyAI | deAPI | Self-hosted |
|---|---|---|---|---|---|---|
| Streaming | ✗ | ✗ | ✓ | ✓ | ✗ | DIY |
| Diarization | ✗ | ✗ | ✓ | ✓ | ✗ | via pyannote |
| Languages | 99 | 99 | 40 | 99 | 99 | 99 |
| URL transcription | ✗ | ✗ | ✗ | ✗ | ✓ | ✗ |
| Max file size | 25 MB | 25 MB | No limit | No limit | 50 MB (URL: no limit) | Your GPU memory |
| Timestamps | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| Translation (→EN) | ✓ | ✓ | ✗ | ✗ | ✓ | ✓ |
| initial_prompt | ✓ | ✓ | ✗ | Word boost | ✓ | ✓ |
| OpenAI SDK compatible | Native | ✓ | ✗ | ✗ | ✓ | ✗ |
| Free tier | None | Rate-limited | $200 credit | $50 + 185 hrs | $5 credit (~237 hrs) | GPU cost |
Two camps emerge. Deepgram and AssemblyAI compete on features - streaming, diarization, audio intelligence built in. OpenAI, Groq, and deAPI compete on Whisper compatibility and simplicity. Self-hosting sits in its own lane: maximum control, minimum hand-holding.
The decision axis is straightforward. Voice assistants and live captions need Deepgram's streaming. Meeting recordings with speaker labels need AssemblyAI. YouTube backlogs and batch workloads need deAPI or Groq at a fraction of the cost.
Cost at scale: 500 hours per month
Abstract pricing means nothing without volume context. Here's what 500 hours of monthly transcription costs on each platform:
| Provider | Monthly cost (500 hrs) |
|---|---|
| OpenAI | $180.00 |
| Deepgram (batch) | $130.00 |
| AssemblyAI (Best) | $375.00 |
| AssemblyAI (Nano) | $60.00 |
| Groq | ~$10.00 |
| deAPI | $10.52 |
| Self-hosted (L4 GPU) | $25-75 (infra) |
The difference between $180/month (OpenAI) and $10.52/month (deAPI) buys you a lot of other API calls.
When Whisper isn't the right model
Whisper excels at batch transcription of clean-to-moderate audio across dozens of languages. It starts falling short in specific scenarios.
Phone and call-center audio recorded at 8 kHz exposes Whisper's weakness. Deepgram Nova-3 was built for this - their WER on telephony audio is 9.4% vs. Whisper's 12.8%. If your audio comes from phone lines, Deepgram or Speechmatics will produce measurably better output.
Real-time voice applications need sub-300ms latency. Whisper is batch-only across every hosted provider. Deepgram's streaming endpoint and AssemblyAI's Universal-Streaming are the viable options here.
Heavy accent and code-switching scenarios - speakers mixing languages mid-sentence - benefit from models trained specifically for that pattern. Speechmatics and Deepgram handle this better than vanilla Whisper.
For everything else - podcast transcription, YouTube content, meeting recordings, multilingual batch processing - Whisper Large V3 through any of the hosted options above will get the job done.
The bottom line
Six providers, one model, wildly different trade-offs.
Real-time streaming or speaker labels rule out every Whisper-based option - go with Deepgram or AssemblyAI. If your input is YouTube, Twitch, or X Spaces URLs, deAPI is the only provider that skips the download-extract-upload pipeline. And if cost drives the decision, deAPI ($0.021/hr) and Groq (~$0.02/hr) run the same model for 17x less than OpenAI.
The transcription quality is comparable across the board. What separates these providers is the engineering you do (or don't have to do) around it.
Prices verified June 2026. All platforms update pricing regularly - check their docs for current rates.
Try deAPI: app.deapi.ai - $5 free credits on signup, no credit card. The /transcribe endpoint accepts YouTube, Twitch, TikTok, Kick, and X URLs directly.
Top comments (0)