Deepgram's own pricing page is charging 38 percent less for its most common transcription rate than it was showing a few months ago. Nova-3 streaming transcription, the speech-to-text model most voice AI builds default to, currently lists at $0.0048 a minute against a struck-through former rate of $0.0077 a minute, with the word "promo" next to it and no date for when that promo stops.
TL;DR:
- Nova-3 streaming transcription on Pay As You Go currently costs $0.0048 a minute monolingual, down from a regular rate of $0.0077, per Deepgram's own pricing page.
- That discount is explicitly labeled promotional, with no expiration date stated anywhere on the page.
- Several competing "Deepgram pricing" roundups still quote the old $0.0077 rate as current, because they were written before the price dropped.
- A $200 free credit with no expiration covers roughly 43,000 minutes, over 700 hours, of Nova-3 transcription before billing starts.
- Buying Deepgram directly is cheaper than getting it bundled inside Vapi's or Retell's per-minute rate, which marks up the same underlying transcription cost.
What Deepgram actually charges right now
Deepgram sells speech-to-text in two shapes: streaming, for live calls where the transcript has to keep up with the conversation, and pre-recorded, for batch processing audio that already exists.
On the Pay As You Go plan, Deepgram's pricing page lists these current Nova-3 rates:
- Streaming, monolingual: $0.0048 a minute
- Streaming, multilingual: $0.0058 a minute
- Pre-recorded, monolingual: $0.0043 a minute
- Pre-recorded, multilingual: $0.0052 a minute
The Growth plan, which requires a prepaid annual credit starting around $4,000, discounts those further to $0.0042, $0.0050, $0.0036, and $0.0043 a minute respectively. Enterprise pricing is custom and scales with volume and deployment requirements.
Billing runs per second of processed audio, not rounded up to the next minute, which matters most for short calls where rounding would otherwise pad the bill.
The part most guides miss: this is promo pricing, with no end date
Here is the detail that does not survive copy-pasting into a roundup post. Next to the current streaming rate, Deepgram's own page shows the previous rate struck through: "$0.0048 (was $0.0077) / min." That is a 38 percent cut on the single most commonly used rate on the page, and the only label attached to it is "promo."
This is not a hypothetical concern. Checking several currently-ranking "Deepgram pricing" pages directly turns up a split: some correctly flag the $0.0048 figure as a limited-time promotional rate against a $0.0077 regular rate, while others state $0.0077 as the current flat price with no mention of a discount at all, likely because they were written or last updated before the cut took effect. Either way, a number copied from a cached table is not the same as a number read off Deepgram's live page the day you price a project.
What the gap looks like at real call volume
The difference between the promo rate and the regular rate is small on a single call and adds up fast at the volume a real deployment runs.
As an illustration: a voice agent handling 500 hours of streaming calls a month (30,000 minutes) costs about $144 a month for transcription at the current $0.0048 promo rate, versus about $231 a month at the $0.0077 regular rate. That is roughly an $87 a month swing on transcription alone, before the language model or text-to-speech voice are added on top. Neither figure is a real client's bill, it is a round number chosen to make the comparison easy to follow, but the ratio between the two holds regardless of volume.
Scale that illustration up or down to your own call volume and the point stays the same: the gap between the two rates is not a rounding error once a deployment runs more than a handful of hours a month. It is the difference between a budget that holds and one that needs a second look the day the promo label disappears.
The free credit and what it actually buys
New Deepgram accounts get $200 in free credit, no credit card required, with no stated expiration date. By Deepgram's own estimate on its pricing page, that covers roughly 43,000 minutes, over 700 hours, of Nova-3 transcription before any billing starts.
Concurrency is the other limit that matters once a project moves past testing. Pay As You Go allows 150 concurrent streaming connections. Growth raises that to 225, though Voice Agent connections specifically are capped lower, at 60 in North America and 45 in Europe, Australia, and India. For most small and mid-size voice AI deployments, the Pay As You Go concurrency ceiling is high enough that the per-minute rate, not the connection limit, decides the real cost.
Why the direct rate is not what most builds actually pay
Deepgram's own rate is the floor, not the ceiling, for what a production voice agent pays for transcription. Most teams do not call Deepgram directly. They buy it bundled inside a platform like Vapi or Retell AI, and that platform marks up the underlying provider cost to cover its own orchestration layer.
Vapi's own pricing page lists Deepgram transcription inside its stack at $0.0095 to $0.0099 a minute. Against Deepgram's current $0.0048 promo rate, that is roughly double. Before the promo cut, when Deepgram's own rate was $0.0077, the same Vapi figure was only about 25 to 30 percent higher, not double. The gap between buying direct and buying bundled did not change, only how large it looks did, since Deepgram's side of the comparison moved and Vapi's did not. That gap is not a markup on quality. It is the cost of not having to build the orchestration, call routing, and model integration yourself. Our breakdown of voice AI pricing across every component in the stack walks through how telephony, the language model, and text-to-speech stack on top of transcription to produce the number that actually shows up on an invoice.
Going direct to Deepgram only makes sense if you are also handling the orchestration layer yourself, which is the trade-off our LiveKit build guide covers in more detail for teams weighing self-hosted infrastructure against a managed platform.
Where this fits inside a bigger build
Transcription is rarely the whole cost of a voice AI project, and it is usually the smallest line item once a language model and text-to-speech voice are added on top. Our retell vs VAPI vs Bland comparison breaks down how those platforms price the rest of the stack, if Deepgram's rate is one piece of a larger build decision rather than the whole question.
If the end goal is a working phone agent, not just a transcription line item, that is the kind of build our voice AI team scopes end to end, pricing the full stack against current vendor rates rather than a cached table. When transcription or any other API call is one step inside a larger pipeline that also writes to a CRM or triggers follow-up actions, that orchestration is what our AI automation team builds around it.
The decision rule
When a Deepgram rate is going into a quote or a budget with more than a few months of runway, price it against the $0.0077 regular rate shown on the page, not the $0.0048 promo rate sitting next to it. If the promo is still running when the bill comes due, the project costs less than planned. If it is not, nothing breaks. Pricing off the number with no stated expiration date is the only version of this math that cannot surprise you later.



