Transcription has the widest price spread of any category I route, and unlike search, the spread is mostly one outlier: Groq hosting Whisper at a price that looks like a typo. $0.00084 per audio minute through our router. Not $0.0084. Under a tenth of a cent per minute, which works out to about five cents per hour of audio.
The obvious question is what you give up for that, and the honest answer is: less than you would think for some workloads, and everything that matters for others. Here are the three transcription providers in our catalog, cheapest first.
1. Groq Whisper large-v3-turbo: $0.00084 per minute
Groq runs OpenAI’s open-weight Whisper large-v3-turbo model on their custom inference hardware, and the resulting price is the floor of the category by a wide margin. Our catalog scores it 80, which is the right way to think about it: Whisper is a genuinely good model, and this is genuinely good transcription, not a bargain-bin compromise.
What you get is the model’s raw competence. What you do not get is the layer of product that managed speech-to-text vendors build on top. The docs suggest the feature set is essentially “audio in, text out”, and in my experience that is exactly how to plan for it. If your pipeline needs speaker labels, aggressive formatting, or streaming, this is not where you get them.
2. Deepgram Nova-3: $0.00516 per minute
Deepgram is a speech company first, and Nova-3 is their flagship model. At $0.00516 per minute routed it costs about 6x Groq, and our catalog scores it 88. The premium buys the managed-STT product surface: Deepgram’s docs lead with features like diarization, smart formatting, and real-time streaming, which is precisely the layer the cheap tier skips. If transcription output goes in front of humans, or downstream code depends on knowing who said what, this tier is where those requirements get met.
3. ElevenLabs Scribe: $0.00804 per minute
Scribe is the quality ceiling of our transcription catalog at a score of 91, and the price ceiling at $0.00804 per minute routed, nearly 10x Groq. ElevenLabs built its reputation on the speech side, and Scribe is their transcription entry: the positioning is accuracy-first. If you are choosing purely on our quality scores, it is the top pick, with the corresponding bill.
The math at real volumes
Per-minute prices are abstract, so here is what a month of audio actually costs. Routed prices, which are provider list plus 20% (the pricing docs cover the math):
| Monthly volume | Groq Whisper | Deepgram Nova-3 | ElevenLabs Scribe |
|---|---|---|---|
| 100 hours | $5.04 | $30.96 | $48.24 |
| 1,000 hours | $50.40 | $309.60 | $482.40 |
| 10,000 hours | $504 | $3,096 | $4,824 |
At small volumes the spread is pocket change and you should just pick on quality. At 10,000 hours a month, the gap between the floor and the ceiling is over $4,000 monthly, and suddenly the question “does this workload actually need the premium tier?” is worth an afternoon of testing.
When cheap wins
The pattern I keep seeing: a large share of agent transcription is machine-to-machine. The text never reaches a human eyeball. It gets embedded for search, summarized by a model, scanned for keywords, or used to decide what an agent does next. For those jobs, Whisper-level accuracy at Groq’s price is not a compromise, it is the correct engineering decision:
- Search indexing: transcribing archives so they are findable. Occasional errors barely dent retrieval.
- Drafts and notes: internal meeting notes, first-pass drafts a human will edit anyway.
- Agent context: audio an LLM will summarize. The summary model papers over small transcription errors.
When to pay up
- Anything compliance-adjacent: if a transcript could end up in a legal or medical context, accuracy is not a cost line, it is the product.
- Captions and publishing: text shown to users carries your brand. Errors are visible and embarrassing.
- Multi-speaker audio where attribution matters: if you need diarization, the cheap tier does not offer it, so the comparison is moot.
- Real-time: streaming transcription is a managed-vendor feature.
Or route by workload
This category is a textbook case for not making one choice. Cheapest-first routing sends the bulk machine-to-machine volume to Groq and keeps failover behind it (failover on our router tries up to 3 providers, and per the routing docs you can flip the default with no code change). Quality-first routing on the human-facing subset lands on Scribe or Nova-3. The per-unit billing that makes this arithmetic possible is something I wrote up in understanding per-unit pricing.
The cheapest transcription API in 2026 is Groq’s Whisper, and for a surprising fraction of real workloads it is also the right one. The full side-by-side, generated from the same open catalog our router reads, is on the transcription API comparison page.