Deepgram pricing: speech API rates and agent cost math

Deepgram pricing at a glance

Deepgram's Standard Voice Agent API costs $0.075 per connected minute on Pay as you go. Standalone Nova-3 Monolingual transcription costs $0.0077 per audio minute for streaming or $0.0043 for prerecorded audio. Those prices buy different services.

The November 2026 rates below are in U.S. dollars. Usage examples show charges before free credits or promotions. Deepgram's pricing separates three main billing units:

Audio Intelligence, including summarization and sentiment analysis, adds a token-based charge to transcription. A token is a unit of text processed or generated by a model.

A transcription minute is not a complete voice-agent minute. Likewise, standalone TTS characters are not another charge to add automatically to the bundled Voice Agent rate.

Plans, free credits and the Growth commitment

Deepgram has two listed self-serve plans and a quote-based Enterprise offer.

The $200 starting credit does not expire, and no card is required to start. It is a dollar balance, rather than a recurring monthly allowance of free minutes. There is no single free-minute allowance to apply across the rate tables.

Growth is an annual prepayment. Dividing $4,000 by 12 gives a planning equivalent of about $333.33 per month, but does not turn it into a monthly subscription.

The advertised discount is up to 20%, not a uniform reduction. For example, Nova-3 Monolingual streaming drops from $0.0077 to $0.0065 per minute, approximately 15.6%. Flux TTS has the same listed rate on both plans.

Lower usage rates also do not erase the commitment. Assume an entire year's workload is 10,000 Standard Voice Agent minutes. Pay as you go usage is $750; Growth usage is $680, but entry to Growth still requires at least $4,000 prepaid. Comparing $750 with $680 alone would hide most of the cash decision.

Growth overages and unused balances

Once Growth credits run out, overage usage costs the Growth rate plus 10%. It is charged weekly on Mondays for the prior week. A card must remain on file; otherwise, requests return HTTP 402 after the credits are exhausted.

For Standard Voice Agent, that makes the overage rate $0.0748 per minute: $0.068 × 1.10. It is different from the $0.068 rate funded by prepaid credits.

Renewing early means purchasing a new Growth plan. This resets credits to the “standard rate” and moves the billing cycle forward. The published terms do not define “standard rate” or explain what happens to unused annual credits, including rollover or expiration at term end. The free starting credit's no-expiry policy does not establish a Growth rollover policy.

Deepgram accepts refund requests for unused credits purchased within the previous 30 days, subject to eligibility review. That is not a guaranteed refund of every unused balance.

Speech-to-text pricing: model, mode and channels

Flux is a streaming speech-recognition family with model-native turn detection for conversational applications. Nova-3 supports both batch and streaming transcription. Their different model capabilities matter even where the per-minute rates match.

Streaming STT rates

These rates are per processed audio minute, before optional add-ons. Feature availability depends on the selected model.

Prerecorded STT rates

Prerecorded processing has a separate rate table. A streaming rate should not be used to price an uploaded recording.

Normal STT audio requests are billed per second, with no block rounding. A 14-second request uses 14 seconds of billable audio duration.

Channels change that duration. A ten-minute stereo recording contains twenty audio minutes across its two channels. At the Nova-3 Monolingual prerecorded rate, its transcription charge is:

• Pay as you go: 10 × 2 × $0.0043 = $0.086.
• Growth: 10 × 2 × $0.0036 = $0.072.

Both calculations exclude optional paid features. Speaker labels and audio channels are separate concepts: diarization identifies speakers, while multichannel billing sums the audio processed across channels.

STT add-ons: diarization differs by mode

Smart Formatting is included. The following charges are additional per audio minute:

This mode distinction changes a real estimate. Assume 60 minutes of mono Nova-3 Monolingual streaming with Speaker Diarization:

• Pay as you go: 60 × ($0.0077 + $0.0020) = $0.582.
• Growth: 60 × ($0.0065 + $0.0017) = $0.492.

Without diarization, the same streaming transcription costs $0.462 or $0.390. For prerecorded audio, adding the streaming diarization surcharge would overstate the charge.

For budgets spanning speech providers, our speech-to-text pricing guide explains how channel counts, billing increments and feature charges affect comparisons.

Text-to-speech pricing: count input characters

Standalone TTS is billed by input text length. Generated speech duration is not the published billing unit.

Assume exactly 10,000 input characters, with no other usage or promotional credit applied. Divide the character count by 1,000, then multiply by the model rate:

A five-minute call does not supply enough information to calculate a standalone TTS bill. It does not tell you how much text the agent generated.

Flux TTS also has a matching-credit promotion: $1 in credits per $1 spent, capped at $500, through December 31, 2026. This is separate from the base rate. The table and examples do not treat a future promotional credit as an immediate discount.

The Voice Agent voice options support the full Flux TTS catalog in English and 41 other languages. Set agent.speak.provider.version to v2 and select a flux-{voice}-{language} model. Those 41 non-English languages are available on api.deepgram.com, api.eu.deepgram.com and api.in.deepgram.com, but not yet on Australia's api.au.deepgram.com endpoint. Languages marked Early Access in the Flux voice catalog do not yet normalize text; numbers, dates and abbreviations need to be written out in words.

Voice Agent API pricing: budget connection time

Standard and Advanced combine STT, a managed large language model (LLM) and TTS in one session. A WebSocket is the persistent connection carrying the session. Its connected duration is the billing clock.

“Bring your own” is often abbreviated BYO. Those rows lower Deepgram's charge by moving components outside its managed bundle. They do not establish the total cost of external providers.

Standard and Advanced have different LLM model assignments. Choosing a model can therefore change the applicable tier, even when the application's call volume stays the same.

Sixty connected minutes, two different bundles

Assume exactly 60 billable connected minutes, with no additional standalone speech requests:

The second row is only the Deepgram portion. External LLM and TTS costs remain outside that subtotal. The first row already includes its managed speech and language-model components, so adding standalone Nova-3 and Aura charges would double-count them.

Neither example includes taxes, phone service or application infrastructure. Deepgram's phone-agent integration connects a separate carrier account to the agent service; the agent rate is not a complete telephone-call budget.

The published connection-time rule does not specify partial-minute rounding, idle-connection exclusions or reconnect treatment. The examples use whole billable connected minutes and assume no speech-only discount. STT's per-second rounding policy cannot automatically be transferred to Voice Agent sessions.

Audio Intelligence adds token charges

Summarization, Topic Detection, Sentiment Analysis and Intent Recognition use the following rates on top of transcription:

Assume one Sentiment Analysis workload with 100,000 input tokens and 10,000 output tokens:

• Pay as you go: (100 × $0.0003) + (10 × $0.0006) = $0.036.
• Growth: (100 × $0.00024) + (10 × $0.00048) = $0.0288.

Those are analysis charges only. Transcription remains a separate line item. This token-based add-on also differs from the per-audio-minute STT features listed earlier.

Concurrency and quote-only costs

Credits fund volume. Concurrency limits how many requests or sessions can run simultaneously. An affordable annual workload can still exceed its peak capacity.

Deepgram's API rate limits vary by service, model, region and plan. For example:

• Pay as you go Voice Agent allows up to 45 concurrent connections in North America, Europe, Australia and India.
• Growth Voice Agent allows up to 60 in North America and 45 in the other three regions.
• Growth Nova-3 streaming allows up to 225 requests in North America, versus 150 in the other three regions.
• An add-on can impose a lower limit. Streaming Speaker Diarization on Pay as you go allows up to 50 requests in North America and 25 in the other regions.

Limits apply per project, rather than per API key. Additional keys do not create another concurrency pool. When a request uses multiple services, the lower service limit applies.

Enterprise rates are quote-only, with no published commitment floor or contract term. The listed self-serve Custom BYO rows are not Enterprise quotes. Dedicated or self-hosted deployments also need separate commercial terms, rather than an extrapolation from the hosted per-minute rates. In a self-hosted deployment, customer infrastructure, backups and monitoring are additional responsibilities described in the deployment responsibility split.

Keep the estimate tied to the architecture

A reproducible Deepgram budget has separate lines for:

1. Standalone STT: processed seconds across channels, model, mode and selected add-ons.
2. Standalone TTS: input characters and model.
3. Voice Agent: connected minutes and the applicable managed or BYO tier.
4. Audio Intelligence: input and output tokens for the selected analysis features.
5. Costs outside those rates: carrier service, external BYO providers, application infrastructure and taxes.

Only include the lines your implementation actually uses. A custom speech stack and a bundled Voice Agent session are different purchasing paths. Our voice-agent pricing guide explains the wider runtime, model and carrier budget.

If you only need transcription or synthesized speech inside an existing application, the relevant Deepgram component rate may be enough for that purchase. If you are choosing the foundation for a production voice AI product, runtime and operating workflows belong in the decision too.

At Dasha, we help technical teams build and run production voice AI agents through a managed runtime, REST APIs and a web application, with telephony, integrations, testing and monitoring. That is a separate platform decision from buying a standalone speech API; it does not imply a like-for-like price or savings against Deepgram.

For a managed production foundation, start your Dasha evaluation.