Transcription has the most interesting price gap in our whole catalog, because both ends of it are defensible. Groq serves Whisper large-v3-turbo at $0.00084 per audio minute routed. Deepgram’s Nova-3 costs $0.00516 per minute routed. That is a 6.1x difference for two APIs that both take audio in and hand text back, and unlike some price spreads I write about, this one is not irrational on either side.
I route transcription traffic across both every day, so let me lay out what the gap actually buys.
The numbers first
Routed prices are provider list plus our 20% margin, per the pricing docs. Quality scores are our hand-curated 0-100 ratings from the open catalog.
| Provider | Model | Routed per audio minute | List price | Quality score |
|---|---|---|---|---|
| Groq | whisper-large-v3-turbo | $0.00084 | $0.0007 | 80 |
| Deepgram | Nova-3 | $0.00516 | $0.0043 | 88 |
| ElevenLabs | Scribe | $0.00804 | $0.0067 | 91 |
Scaled up, the gap gets vivid. An hour of audio costs about $0.05 through Groq and about $0.31 through Deepgram. A thousand hours, which is a real number if you are backfilling a podcast archive or indexing a call library, runs roughly $50 versus $310. I included ElevenLabs Scribe in the table because it sits above both and matters for the quality ceiling, but this post is about the two poles most people are actually choosing between.
What Groq is selling
Groq is not selling a transcription product. It is selling extremely cheap, extremely fast inference on an open model. Whisper large-v3-turbo is a known quantity: the weights are public, its behavior is well documented across languages and accents, and the transcript you get from Groq is the same class of output you would get self-hosting the model, minus the ops work.
That framing explains both the price and the limits. You get audio in, text out, at a price that rounds to free for most workloads. What you do not get is the managed layer that a dedicated speech company builds around its model: no speaker labels, minimal control over formatting, and none of the vocabulary-boosting knobs that matter when your audio is full of product names and jargon. The 80 quality score in our catalog reflects that: strong raw transcription, thin product surface.
What Deepgram is selling
Deepgram is a speech company, and Nova-3 is its flagship model wrapped in an actual product. The docs advertise diarization (who said what), smart formatting (dates, currency, phone numbers rendered the way a human would write them), keyterm prompting to boost recognition of domain vocabulary, and streaming support for live audio. I have not benchmarked raw word error rates between the two models, and I will not invent numbers here, but the 88 score we assign Nova-3 is mostly about that feature depth and output polish. The methodology behind those scores is written up at /docs/quality if you want to know exactly what they do and do not claim.
The honest summary: you are not paying 6x for 6x fewer errors. You are paying 6x for the parts of transcription that happen after the words are recognized.
When the cheap option wins
More often than people expect. If the transcript is an intermediate artifact, meaning it gets fed to an LLM for summarization, chunked into a search index, or skimmed by one engineer, Whisper via Groq is the right call. Downstream models are remarkably tolerant of small transcription errors, and nobody is reading the raw text anyway. This is the workload I covered in my cheapest-transcription piece: drafts, backfills, and indexing pipelines, where a 6x price difference compounds into real money and the feature gap costs you nothing.
When Deepgram earns the gap
The moment a human reads the transcript, the calculus flips. Meeting notes without speaker labels are barely usable. Captions with mangled number formatting look broken. Support-call analytics where the model keeps missing your product’s name will quietly rot your data. Those are exactly the problems Deepgram’s managed features exist to solve, and at $0.31 per hour it is still cheap in absolute terms. If accuracy itself is the entire point, say compliance or legal contexts, that is where I would look one rung higher at Scribe.
How I actually route it
My default is cheapest-first: Groq takes the traffic, and if it errors or times out, failover walks up the chain (the router tries up to three providers per request). For anything user-facing I flip to quality-sorted routing, which is a one-line preference change per the routing docs rather than a code change. The two-tier structure of this category makes it unusually well suited to that split: cheap tier for machine consumers, managed tier for human ones.
Full pricing for every transcription provider we route, generated from the same catalog, is on the transcription API comparison page.