Deepgram vs AssemblyAI vs Google Speech-to-Text: Real Transcription Cost per 10,000 Hours in 2026
The short answer: the rate-card leader on paper is not necessarily the cheapest for your workload, because diarization, redaction, and streaming versus batch each carry their own surcharge, and none of the three vendors here bundles a feature the way it first appears to. At 10,000 hours of clean English audio, AssemblyAI's Universal-2 batch model costs about $1,500, while Deepgram's Nova-3 batch is about $4,620 before any add-ons and Google's Dynamic Batch tier is about $2,100. On multilingual, streaming meeting audio that needs diarization, AssemblyAI's realtime model runs about $5,700 ($0.45/hour base plus $0.12/hour for realtime diarization, itself pricier than the pre-recorded diarization add-on), against Deepgram's $5,820 and Google's $9,600. Deepgram and AssemblyAI bill speaker diarization as a separate add-on; Google's diarization pricing was not consistently itemized in the sources reviewed and remains QUOTE-ONLY / UNKNOWN in this comparison. Comparing a rate that omits diarization to one that includes it is the single most common way to misprice this category.
What each vendor bills
Deepgram (REPORTED, deepgram.com/pricing as read by multiple third parties through August–September 2026). Nova-3 pay-as-you-go: $0.0077 per minute for both streaming and pre-recorded/batch transcription (about $0.46 per hour). A separate, apparently promotional or older figure of $0.0048 per minute for "monolingual streaming" appears on other pages, so the two should be reconciled against the current live pricing page before budgeting; this article uses the higher, more consistently sourced $0.0077 figure. Add-ons meter separately: speaker diarization +$0.0020/min, redaction +$0.0020/min (ILLUSTRATIVE parity with diarization; exact redaction rate was not independently confirmed), keyterm prompting +$0.0013/min. Stacking diarization and redaction together adds roughly 52% to the base rate. A $200 free credit buys roughly 430 hours of transcription. The Growth plan requires a minimum annual spend of $4,000 and lowers the effective rate to about $0.0065 per minute, a savings of about $72 per 1,000 hours relative to pay-as-you-go, alongside higher WebSocket concurrency.
AssemblyAI (OFFICIAL, AssemblyAI's current pricing reference). Universal-2 async/batch base: $0.15 per audio hour ($0.0025/min). Universal-3.5 Pro async base: $0.21/hour ($0.0035/min). Universal-3.5 Pro Realtime base: $0.45/hour ($0.0075/min), roughly 2.1 times the async rate for the same model tier because streaming carries its own premium. Speaker diarization is a separate, itemized add-on on every AssemblyAI product, not included in any base rate: pre-recorded (async/batch) diarization is +$0.02/hour, and realtime diarization is +$0.12/hour — six times more expensive than the pre-recorded version, because diarizing a live stream is a harder problem than diarizing a completed recording. Redaction is likewise separate: PII Audio Redaction (bleep or replace spoken PII) is +$0.05/hour, and PII Text Redaction (redact the transcript) is +$0.08/hour. Profanity filtering (+$0.01/hour) and content moderation (+$0.15/hour) are separate add-ons again. A July 1, 2026 pricing change was noted by one source without full details captured (QUOTE-ONLY / UNKNOWN for the specifics of that change). A $50 free credit and a 100-minute monthly free tier (batch only) are available.
Google Cloud Speech-to-Text (REPORTED). Standard V2 streaming recognition: from $0.016/min ($0.96/hour). Dynamic Batch (Chirp) for non-urgent, 24-hour-turnaround transcription: roughly $0.003–$0.004/min ($0.21–$0.24/hour), the cheapest tier Google offers by a wide margin, but only for workloads that can tolerate delayed processing. 60 minutes a month are free, with no credit card required for the first 12 months. Diarization pricing on Google was not consistently itemized in the sources reviewed (QUOTE-ONLY / UNKNOWN); channel handling also differs from other vendors, with Google billing multi-channel audio separately in some configurations.
Three audio shapes
All ILLUSTRATIVE. A: clean English audio — standard batch transcription, no diarization needed. B: multilingual meetings — streaming, with diarization (who-said-what) required for a multi-speaker recording. C: noisy call-center audio with diarization — batch processing, diarization plus redaction (removing sensitive spoken data like card numbers) for compliance.
Cost per 1,000 and 10,000 hours by shape
Formula: cost = hours × 60 × rate per minute, with add-ons summed where the shape requires them.
Shape A: clean English, batch, no diarization
| Hours | Deepgram (Nova-3 batch) | AssemblyAI (Universal-2, no diarization needed) | Google (Dynamic Batch) |
|---|---|---|---|
| 100 | $46.20 | $15.00 | $21.00 |
| 1,000 | $462.00 | $150.00 | $210.00 |
| 10,000 | $4,620.00 | $1,500.00 | $2,100.00 |
Shape B: multilingual meetings, streaming + diarization
Formula: AssemblyAI = Universal-3.5 Pro Realtime base ($0.45/hour) + realtime diarization (+$0.12/hour) = $0.57/hour.
| Hours | Deepgram (Nova-3 stream + diarization) | AssemblyAI (Universal-3.5 Pro Realtime + diarization) | Google (standard streaming) |
|---|---|---|---|
| 100 | $58.20 | $57.00 | $96.00 |
| 1,000 | $582.00 | $570.00 | $960.00 |
| 10,000 | $5,820.00 | $5,700.00 | $9,600.00 |
Shape C: noisy call-center audio, batch + diarization + redaction
Formula: AssemblyAI = Universal-3.5 Pro async base ($0.21/hour) + pre-recorded diarization (+$0.02/hour) + PII Audio Redaction (+$0.05/hour) = $0.28/hour.
| Hours | Deepgram (Nova-3 batch + diarization + redaction) | AssemblyAI (Universal-3.5 Pro async + diarization + PII Audio Redaction) | Google (Dynamic Batch, diarization rate not itemized) |
|---|---|---|---|
| 100 | $70.20 | $28.00 | $21.00 (base rate only; diarization not separately priced in sources reviewed) |
| 1,000 | $702.00 | $280.00 | $210.00 (base rate only) |
| 10,000 | $7,020.00 | $2,800.00 | $2,100.00 (base rate only) |
Adding PII Text Redaction as well, for cases needing both audio bleeping and a redacted transcript ($0.21 + $0.02 + $0.05 + $0.08 = $0.36/hour), raises the AssemblyAI figure to $36, $360 and $3,600 at the three volumes. The Google figures on shape C understate the true cost, because Google's diarization surcharge was not itemized in the sources reviewed; treat the Google column on shape C as a floor, not a complete quote.
Premium-feature uplift
Formula: uplift % = add-on cost ÷ base rate. On Deepgram, diarization alone adds 26% to the Nova-3 batch rate, and diarization plus redaction together add roughly 52%. On AssemblyAI, every add-on is itemized separately and none is bundled: pre-recorded diarization adds about 10% to the Universal-3.5 Pro async rate ($0.02 on $0.21), while realtime diarization adds about 27% to the Universal-3.5 Pro Realtime rate ($0.12 on $0.45) — a much steeper uplift, reflecting that realtime diarization is priced at six times the pre-recorded add-on's dollar rate. PII Audio Redaction alone adds about 24% to the Universal-3.5 Pro async rate ($0.05 on $0.21), and stacking pre-recorded diarization plus both redaction types brings the combined uplift on that same base to about 71% ($0.15 of add-ons on $0.21 of base).
Streaming versus batch
On AssemblyAI, moving from async (batch) to realtime (streaming) on the same Universal-3.5 Pro model multiplies the base rate by about 2.1x ($0.0035 to $0.0075 per minute), and diarization itself gets more expensive too: six times as much per hour in realtime as in batch. On Deepgram, the pay-as-you-go rate reviewed here is stated as the same $0.0077 for both streaming and batch, so Deepgram shows no streaming premium on its base rate in this comparison, in contrast to AssemblyAI's compounding streaming surcharge on both the base rate and the diarization add-on. Google's gap is the widest on the base rate alone: standard streaming at $0.016/min versus Dynamic Batch at roughly $0.0035/min is a 4.6x difference, reflecting that Dynamic Batch trades a 24-hour turnaround for a steep discount.
Sensitivity
- Which Deepgram rate applies. A $0.0048 versus $0.0077 per-minute conflict across sources is a 60% swing; confirm the current live rate before committing to volume.
- Committed-use discounts. Deepgram's Growth plan requires a $4,000 annual minimum for a roughly 16% rate reduction; below that spend level, pay-as-you-go is the only option.
- Realtime versus pre-recorded diarization on AssemblyAI. The same feature costs six times more per hour in realtime ($0.12) than in batch ($0.02); a workload that can tolerate a short processing delay and use the async product instead of realtime saves substantially on diarization alone, separate from the base-rate saving.
- Turnaround tolerance. A workload that can wait 24 hours unlocks Google's Dynamic Batch tier at roughly a quarter of its standard rate.
Budgeting traps
- Assuming any vendor's diarization is bundled. Do not assume diarization is bundled into the base rate. Deepgram and AssemblyAI price it separately; Google's exact diarization charge was not consistently itemized in the sources reviewed.
- Pricing streaming needs from a batch rate card, or vice versa. The two can differ by more than 2x on the same vendor, and add-on prices can diverge even further between the two modes.
- Assuming Google's Dynamic Batch tier fits a live application. Its steep discount comes with a 24-hour processing delay that rules it out for anything time-sensitive.
- Stacking add-ons without checking the combined rate. Diarization plus redaction plus keyterm prompting on Deepgram, or diarization plus both redaction types on AssemblyAI, can each add 50% or more to the sticker rate.
What to ask before you buy
Ask each vendor to quote your exact feature set (diarization, redaction, language count, streaming versus batch) as one all-in per-minute number, not a base rate plus a list of possible add-ons. Then confirm whether any committed-use discount tier applies at your actual monthly volume.
ARTICLE 13