Transcription has the largest price ratio of any category I route: the cheapest provider in our catalog costs $0.00084 per audio minute and the most expensive costs $0.00804. That is nearly 10x, within one category, for products that all accept audio and return words. When a spread gets that wide, the interesting question stops being “which is best” and becomes “which failures can you afford.”
Here is the catalog, routed prices (provider list plus 20%):
| Provider | Price per audio minute | Quality score |
|---|---|---|
| Groq Whisper large-v3-turbo | $0.00084 | 80 |
| Deepgram Nova-3 | $0.00516 | 88 |
| ElevenLabs Scribe | $0.00804 | 91 |
What the gap actually costs
Per-minute prices are abstract, so let me multiply. At 10,000 minutes a month (a modest meeting-transcription workload), Groq costs $8.40, Nova-3 costs $51.60, and Scribe costs $80.40. At 100,000 minutes, that is $84 versus $516 versus $804.
Notice what those numbers are not: big. Even the “expensive” option at serious volume is hundreds of dollars, not thousands. This matters for the decision framework, because the argument for the cheap tier is only compelling when volume is high or margins are thin. If you transcribe 2,000 minutes a month, the total gap between floor and ceiling is about $14, and optimizing it is a waste of an afternoon.
When the 10x-cheaper option is fine
Groq’s hosted Whisper at $0.00084 per minute, quality score 80, is one of the best bargains in our entire catalog, and there is a large class of workloads where it is simply the correct answer:
- Drafts a human will edit. If every transcript gets reviewed anyway, errors cost seconds of editing, not trust.
- Search indexing. Embedding transcripts so users can find “the call where we discussed pricing” tolerates a lot of word-level error. Retrieval works on gist.
- Agent context. When the transcript is intermediate material an LLM will summarize, occasional mistranscriptions usually wash out in the summary.
- High-volume archival. Transcribing a backlog you will mostly never read is exactly where a 10x price gap dominates every other consideration.
The common thread: the transcript is an input to another process, not the deliverable. I went deeper on this tier in the cheapest transcription API piece.
When to pay up
The premium tier earns its multiple when the transcript itself is the product, or when errors carry liability:
- Captions and subtitles. Errors ship to every viewer, permanently, and accessibility obligations raise the stakes beyond aesthetics.
- Compliance and legal records. If a transcript might be evidence or a regulatory artifact, “quality score 80” is not a phrase you want in the postmortem.
- Speaker-attributed output. Managed services like Deepgram earn part of their premium with features beyond raw word accuracy: diarization, formatting, and the plumbing around the transcript. Raw hosted Whisper gives you words; turning words into “who said what, punctuated” is engineering you either buy or build. The Deepgram vs Groq Whisper comparison covers that feature gap in detail.
- Customer-visible transcripts. Anything a user reads under your logo. Every error is a tiny credibility tax.
Note that quality scores here (80 vs 88 vs 91) are our curated ratings, not word-error-rate benchmarks. The gap is real but it is a band, and audio conditions (accents, crosstalk, noise) move outcomes more than provider choice at the margin. Bad audio is bad audio at every price.
A decision rule that fits on a sticky note
Ask two questions. Does a human read this transcript as a final artifact? Does an error cost more than embarrassment? Two nos: route cheapest, take the 10x savings, spend the difference on literally anything else. Either yes: pay for the 88-or-better tier and sleep.
The refinement most teams miss is that the answer varies per request, not per company. The same product often has both workloads: archival backfill (cheap tier) and live captions (premium tier) in one codebase. This is a routing decision, not a procurement decision. On route.tools you can send each request down a different path with a per-request routing preference, or set saved defaults per workload, and failover means a Groq hiccup lands on Nova-3 automatically rather than dropping the job. Every response tells you the exact price paid, which keeps the accuracy-versus-price ledger honest.
All three providers, priced side by side from the open catalog, are on the transcription comparison page.