Deepgram pricing

About

Deepgram is a speech-AI platform that sells four developer-facing API products: Speech-to-Text (the Nova and Flux model families), Text-to-Speech (Aura), a real-time Voice Agent API, and Audio Intelligence (summarization, topic detection, sentiment, intent). A Y Combinator (W16) company, it serves developers and startups self-serve while landing larger contact-center, conversational-AI, and medical-transcription workloads through Enterprise.

Deepgram raised a $130M Series C in January 2026 at a $1.3B valuation led by AVP, reaching unicorn status; the company reported roughly $21.8M of 2024 revenue, said it turned cash-flow positive in 2025, and states that more than 1,300 organizations build voice AI on its APIs (TechCrunch, 2026-01-13; Deepgram press release). Alongside the raise it acquired OfOne, a voice-AI platform for quick-service-restaurant drive-thru ordering. It competes most directly with AssemblyAI and Google/AWS speech APIs on the transcription side and with Cartesia, ElevenLabs, and OpenAI on the voice/synthesis side.

Unlike seat-priced SaaS, Deepgram meters everything by the work performed: audio minutes for transcription and voice agents, generated characters for synthesized speech, and input/output tokens for audio understanding. This puts it in the same usage-priced cohort as other AI inference providers, where the value metric maps directly to API throughput rather than headcount. See AssemblyAI’s blueprint and Cartesia’s blueprint for adjacent speech-AI pricing models.

For the most current rates, see the Deepgram pricing page.

Pricing summary : per-minute, per-character, and per-token usage

Deepgram runs a pure usage-based pricing model with three distinct billing units across its products, plus a freemium on-ramp and a prepaid discount tier built on credit-based billing. There is no flat seat subscription — you pay only for what you process, and the published per-unit rate depends on which model and which plan column you fall in. See understanding usage-based pricing models for the broader category.

The model has these dimensions:

• Speech-to-Text — per minute of audio. Rate varies by model (Flux, Nova-3 Monolingual/Multilingual), by stream type (Streaming vs Pre-Recorded), and you can view rates as $/minute or $/hour. Example: Nova-3 Monolingual streaming is $0.0048/min on Pay-As-You-Go.
• Text-to-Speech (Aura) — per 1,000 characters. Aura-2 is $0.030/1k characters on Pay-As-You-Go; Aura-1 is $0.0150/1k characters.
• Voice Agent API — per minute. Tiered from $0.075/min (Standard) to $0.163/min (Advanced), with cheaper bring-your-own LLM/TTS variants.
• Audio Intelligence — per 1,000 tokens. $0.0003/1k input and $0.0006/1k output tokens on Pay-As-You-Go (rates opt into the Model Improvement Program).
• STT add-ons — per minute. Redaction and Speaker Diarization each add $0.0020/min; Keyterm Prompting $0.0013/min; Smart Formatting is included free.
• Plan column — Pay-As-You-Go vs Growth. Growth prepays annual credits (from $4K+/year) redeemed against usage at rates up to ~20% lower than Pay-As-You-Go. Enterprise is custom-quoted. See understanding prepaid credits models for how this prepaid-commit pattern works.

What makes this different: Deepgram exposes three different value metrics on a single pricing page — minutes, characters, and tokens — because the unit that best tracks cost differs per product (audio length for transcription, output length for synthesis, payload size for understanding), rather than forcing every product onto one metric.

Pricing by product

All rates below are verbatim from deepgram.com/pricing in USD. The two price columns are Pay-As-You-Go (PAYG) and Growth (prepaid annual credits, up to ~20% cheaper). Streaming Speech-to-Text rates are currently flagged “limited-time promotional”; where the page shows a struck-through original next to the current rate, the original is noted in parentheses.

Plans (account-level)

Speech-to-Text — per minute (Streaming, $/min)

The pricing page also exposes a Pre-Recorded toggle and a $/hour view; rates above are the Streaming, $/minute view captured 2026-05-31.

Speech-to-Text add-ons — per minute (Streaming, $/min)

Text-to-Speech (Aura) — per 1,000 characters

Voice Agent API — per minute ($/min)

Audio Intelligence — per 1,000 tokens

Audio Intelligence rates opt in to the Model Improvement Program.

Sales motions across products: PLG / self-serve for Pay-As-You-Go and Growth (credit-card and prepaid, no quote required); sales-led for Enterprise, Custom STT models, and on-prem/self-hosted deployment.

Hidden costs : What Deepgram users actually pay

Because Deepgram is pure usage, there is no base plan to anchor a bill — your invoice is the sum of per-unit rates times volume across whichever products you touch. The “hidden” costs are the metering-unit switches (minutes vs characters vs tokens), the per-minute add-ons that stack on top of the base STT rate, and the fact that streaming STT currently carries a limited-time promotional rate that will revert to the struck-through original (Nova-3 Monolingual reverts from $0.0048 to $0.0077/min — a 60% increase) once the promo ends. All figures below use Pay-As-You-Go rates captured 2026-05-31.

Archetype A — a transcription startup processing 100,000 minutes/month of streaming Nova-3 Monolingual, with diarization and redaction on.

The add-ons here add $400 — they nearly match the base transcription cost, so a team modeling only the headline $0.0048/min would under-budget by ~83%. When the streaming promo ends, the base line alone jumps from $480 to $770.

Archetype B — a voice-agent product running 50,000 minutes/month of conversation on the managed Standard tier, plus 2M characters of Aura-2 TTS for non-agent flows.

The Voice Agent line dominates, and switching to a BYO-TTS or BYO-LLM tier (Standard – BYO TTS at $0.065/min) would cut $500/month — so the bundling choice, not raw volume, is the biggest lever. At $3,810/month a customer is well past the $4K/year Growth break-even and should prepay for the ~20% discount.

Want to estimate your own Deepgram bill? Use the Deepgram pricing calculator to model your monthly cost based on audio minutes, characters synthesized, model selection, and add-on mix.

Pricing evolution : Deepgram pricing history and changes

Deepgram’s metering model has been consistent for years — per minute of audio for transcription — but the SKU lineup and the rate card have moved as new model families (Nova-3, Aura-2, Flux) and the Voice Agent API shipped. The cadence below is reconstructed from Deepgram’s own product announcements and press coverage; per-snapshot historical screenshots are not yet attached.

Cadence

Tracked range: 2025 Q1–2026 Q2. Dates reflect public product announcements and the live pricing page captured 2026-05-31; quarters not listed had no announced rate or SKU change found. Wayback-backed snapshot dates to be attached on a later pass.

Notable changes

• 2026-05-31 — Live pricing page shows limited-time promotional streaming STT rates: Nova-3 Monolingual at $0.0048/min struck through against $0.0077/min original (captured deepgram.com/pricing). The $0.0077 PAYG / $0.0065 Growth originals match pre-promo rates reported by third-party trackers, corroborating the promo as a discount rather than a price cut.
• 2026-01-13 — Deepgram raised a $130M Series C at a $1.3B valuation and acquired OfOne; pure-usage metering and the published rate card were unchanged by the round (TechCrunch; Deepgram press release).
• 2025 — Nova-3 (STT), Aura-2 (TTS), Flux (streaming STT), and the Voice Agent API shipped, broadening the rate card across three metering units while keeping per-minute as the anchor for audio.

What’s unique : Deepgram’s distinctive pricing mechanics

1. Three metering units on one pricing page. Deepgram doesn’t force every product onto a single value metric. Speech-to-Text and the Voice Agent API bill per minute of audio; Text-to-Speech (Aura) bills per 1,000 characters; Audio Intelligence bills per 1,000 input/output tokens. Each unit is the one that best tracks the underlying cost — audio length for recognition, output length for synthesis, payload size for understanding — which makes per-unit cost legible to buyers but means a customer touching all four products reconciles three different meters on one invoice.

2. A toggle-driven rate card. The pricing page lets buyers re-express the same STT rates as Streaming vs Pre-Recorded and as $/minute vs $/hour. Rather than publishing four separate tables, Deepgram surfaces one rate card and lets the buyer choose the unit that matches how they think about their workload — unusual transparency for an API where most competitors publish a single fixed denomination.

3. Bring-your-own components as discount tiers. The Voice Agent API exposes BYO-LLM and BYO-TTS variants priced below the fully-managed tiers (Standard – BYO TTS at $0.065/min vs $0.075/min managed; Custom – BYO LLM + TTS at $0.050/min). Customers who already pay for their own model or voice stack pay Deepgram only for the orchestration, turning an architecture choice into an explicit pricing lever.

4. Prepaid credits as the only “subscription.” There is no seat or platform fee anywhere. The Growth plan is purely a prepaid credit pool ($4K+/year) redeemed against actual usage at ~20% lower per-unit rates — so the discount is structurally a credit-based volume commitment, not a feature paywall. Free-tier capacity, concurrency, and compliance are the only things that step up by plan.

5. Promotional rates shown as strike-throughs, not silent cuts. The page displays current streaming STT prices struck through against their originals (e.g. $0.0048 over $0.0077), explicitly framing the discount as limited-time. This sets an anchor — buyers see the “real” rate they’ll revert to — which is rare in API pricing, where temporary discounts usually appear as the only number.

Strengths & weaknesses

Billing UX : prepaid credits, auto-load, and refund controls

The pricing page documents a Pay-As-You-Go credit system with explicit account-control mechanics:

• $200 free credit on signup — no credit card required to start; usage draws down the credit before any charge.
• Auto-Load Functionality — the account can automatically top up credits when the balance runs low, so production traffic doesn’t stall.
• Managing Your Subscription — self-serve controls to manage the Growth prepaid-credit subscription.
• Refunds for Unused Credits with a defined Refund Request Process — unused prepaid credits can be requested back per Deepgram’s stated policy.
• Credit Transfer Policy — credits can be transferred between accounts under stated conditions.
• Streaming vs Pre-Recorded and $/minute vs $/hour toggles — the pricing page lets buyers view the same rates under whichever unit matches their workload, improving cost transparency before purchase.
• Concurrency limits surfaced per plan — REST/WSS concurrency ceilings (e.g. up to 150 WSS on PAYG, 225 on Growth for STT) are published so teams can size before hitting throttles.
• Compliance posture published inline — SOC 2 Type 1 & 2, HIPAA (BAA for Enterprise ePHI), GDPR/EU data residency via api.eu.deepgram.com, CCPA, and PCI are listed as billing/trust controls.

Strategic wins : Why Deepgram’s pricing decisions worked

1. Matching the meter to each product’s true cost driver

By billing transcription per minute, synthesis per character, and understanding per token, Deepgram keeps each product’s price legible against the thing that actually drives its cost. This avoids the trap of a single forced metric that over- or under-charges whole product lines. It is a textbook application of choosing the right usage metric — the company accepts invoice complexity in exchange for per-product fairness, which matters when buyers benchmark each API against a specialist competitor.

2. A no-card, no-minimum free credit as the PLG wedge

The $200 free credit with no card and no expiration lets a developer integrate, test at real volume, and ship before any procurement conversation. For an infrastructure API this is the highest-leverage acquisition decision: it converts the evaluation from a sales motion into a self-serve one, exactly the shift away from per-user licenses that defines modern AI infrastructure go-to-market.

3. Prepaid credits as the discount mechanism instead of feature gating

Deepgram unlocks its ~20% discount through a $4K+/year prepaid credit commit (Growth) rather than by paywalling capabilities. Every plan keeps full model access; only price, concurrency, and support step up. This keeps the product undivided and turns the discount into a cash-flow and forecasting win for Deepgram — a clean example of prepaid credit models used to convert usage into committed revenue.

4. BYO tiers that monetize orchestration, not just inference

Offering BYO-LLM and BYO-TTS Voice Agent tiers lets Deepgram capture customers who already own part of the stack, charging them for orchestration alone at a lower per-minute rate. Rather than losing those buyers entirely, Deepgram prices the slice it uniquely provides — a smart response to the margin pressure every voice-AI vendor faces when LLM and TTS costs are commoditizing.

Areas to improve : Gaps in Deepgram’s pricing approach

1. Surface a blended cost-per-conversation estimator

The three-meter model is fair but hard to reason about: a buyer running a voice agent that also transcribes, redacts, and summarizes must hand-assemble four rate lines. Deepgram should publish a worked “cost per 1,000 conversations” example or an interactive estimator on the pricing page itself, so buyers see the all-in number rather than discovering it on the first invoice. This directly addresses the bill-shock and cost-unpredictability risk that pure-usage models create.

2. Make the promotional-rate expiry explicit

Streaming STT rates are flagged “limited-time promotional” with struck-through originals, but no end date is shown. A customer modeling a 12-month budget at $0.0048/min could see a 60% jump to $0.0077/min mid-contract with no warning. Publishing the revert date — or guaranteeing the promo rate for prepaid Growth customers through their credit term — would remove a real forecasting hazard and is consistent with good threshold and alerting practice.

3. Offer a volume-discount step below the $4K Growth commit

Today the only path to discounted rates is a $4K+/year prepaid commit. A mid-volume customer spending $1–2K/year on pay-as-you-go gets no break, creating a cliff. An automatic tiered discount on monthly usage (à la cloud egress tiers) would reward growing accounts before they’re ready to prepay, smoothing the jump from self-serve to committed and reducing churn at the boundary.

Key takeaways

1. Use the metric each product deserves, not one metric for the catalog. Deepgram bills minutes, characters, and tokens because each tracks a different cost driver. If your products have genuinely different cost structures, a single forced value metric will systematically mis-price at least one of them.
2. A no-card free credit is a pricing decision, not a marketing one. The $200, no-minimum, no-expiration credit is what makes the self-serve motion work. For infrastructure APIs, removing the card requirement at signup converts evaluation into adoption far more reliably than a feature-limited free tier.
3. Discount through commitment, not through feature gating. Growth keeps full model access and discounts on price alone via a prepaid commit. Splitting capability across plans fragments your product; discounting on committed volume keeps it whole and converts usage into forecastable revenue.
4. Anchor temporary discounts with the strike-through. Showing the original rate next to a promotional one sets the buyer’s expectation of the durable price — but only works honestly if you also publish when the promo ends. The anchor cuts both ways.
5. Price the slice you uniquely own. BYO-LLM/BYO-TTS tiers let Deepgram keep customers who already own part of the stack by charging only for orchestration. When inputs commoditize, repackage your price around the component competitors can’t replicate.

UBP implications

1. Multi-metric usage pricing is viable when each meter is intuitive. Deepgram proves a single vendor can run minutes, characters, and tokens side by side without confusing buyers — provided each unit obviously matches its product. The cost is invoice reconciliation complexity, which the vendor must offset with estimators and worked examples.
2. Prepaid credits are the cleanest bridge from pure-usage to committed revenue. A pure pay-as-you-go business has no contracted floor; Deepgram’s Growth credit pool converts heavy users into annual prepay without introducing a seat. For UBP companies that need revenue predictability, a credit commit is less disruptive than bolting on a subscription.
3. Promotional rate-card transparency is an emerging UBP norm. Showing struck-through originals signals where prices will land and pre-empts bill-shock backlash. As usage pricing matures, buyers will increasingly expect this anchor — and the discipline of publishing expiry dates alongside it.

Sources

• Deepgram pricing page (accessed 2026-05-31)
• Deepgram documentation (accessed 2026-05-31)
• Deepgram blog (accessed 2026-05-31)
• Deepgram changelog (accessed 2026-05-31)
• Deepgram Series C press release (accessed 2026-05-31)
• Browse the full pricing blueprint for comparable usage-priced AI APIs (accessed 2026-05-31)

Bottom line

Deepgram is a rare example of a pure-usage AI API that stays fully transparent at scale: every model, tier, and add-on has a published rate, three honest meters track three different cost drivers, and the only “subscription” is a prepaid credit pool that discounts price without gating capability. The open questions are forecasting — the unbounded sum of minutes, characters, and tokens — and the limited-time streaming promo whose expiry isn’t dated. For most developers, the $200 free credit means you can find out exactly what you’ll pay before you commit a cent.

Want to compare Deepgram against other speech and voice AI vendors? Browse the pricing blueprint.

Pricing timeline : Major events on a vertical axis

Each milestone below corresponds to a public pricing change, product launch, or material adjustment. Major events use a filled marker; minor adjustments use a faded one.

Facts captured from live pricing page

May 2026

Per-unit rates verified from deepgram.com/pricing: STT per-minute (Flux, Nova-3), TTS Aura per-1k-characters, Voice Agent per-minute, Audio Intelligence per-1k-tokens. Pay-As-You-Go vs prepaid Growth (up to 20% off) columns; $200 free credit; Enterprise sales-led. Limited-time promotional streaming rates in effect.

Blueprint stub created

May 2026

Initial stub generated for batch blueprint expansion. Full pricing history to be populated by wiki-research.

Flux streaming STT model; Streaming/Pre-recorded view split

Dec 2025

STT table split into Streaming and Pre-Recorded toggles and added the Flux model at $0.0077/min streaming. Streaming rates: Nova-3 Monolingual $0.0077/min, Nova-3 Multilingual $0.0092/min, Nova-1 & 2 $0.0058/min, Enhanced $0.0165/min, Base $0.0145/min. Add-ons: Redaction $0.0020/min, Speaker Diarization $0.0020/min, Keyterm Prompting $0.0013/min.

Voice Agent re-priced per minute with BYO tiers; Enterprise becomes Custom Pricing

Aug 2025

Voice Agent API switched from per-hour to per-minute and added a Growth column plus bring-your-own tiers: Standard $0.0800/min PAYG ($0.0700 Growth), Standard-BYO TTS $0.0600/min, Custom-BYO LLM $0.0700/min, Custom-BYO LLM+TTS $0.0500/min, Advanced $0.1600/min ($0.1500 Growth). Enterprise plan card renamed 'Custom Pricing'.

Nova-3 and Aura-2 launch; Voice Agent API debuts (per hour)

Jun 2025

STT adds Nova-3 (English) $0.0043/min PAYG / $0.0036 Growth and Nova-3 (Multilingual) $0.0052/min / $0.0044. TTS adds Aura-2 $0.030/1k characters PAYG ($0.027 Growth); Aura-1 stays $0.0150/1k. New Voice Agent API billed per hour: Standard Tier $4.50/hr PAYG, Standard - Custom LLM $3.90/hr.

Per-product STT / TTS / Audio Intelligence tables; $4k+ / $10k+ plans

Dec 2024

Plans: Pay As You Go (Free + $200 credit), Growth ($4k+/yr, save up to 20%), Enterprise ($10k+/yr). STT pre-recorded — Nova-2 $0.0043/min PAYG / $0.0036 Growth, Enhanced $0.0145/min, Base $0.0125/min, Whisper Large $0.0048/min. TTS Aura $0.0150/1k characters PAYG ($0.0135 Growth). Audio Intelligence $0.0003/1k input - $0.0006/1k output tokens PAYG.

Nova-2 launches; Premium renamed Enterprise

Sep 2023

'Introducing Nova-2: The Fastest, Most Accurate Speech-to-Text API' banner. Three plans now Pay-as-You-Go ($200 free credit, no expiration), Growth (Save ~20% with pre-paid credits, redeemed against actual usage) and Enterprise (custom). Marketed around domain-specific language models (DSLMs).

Nova launches; free credit raised to $200; Growth goes monthly

May 2023

'Introducing Nova: the world's most powerful speech-to-text model' banner. Plans simplified to Pay As You Go ($200 free credit, then pay-as-you-go, no expiration), Growth (Save ~20% with pre-paid credits, $350+/mo) and Premium (custom). Free credit raised from $150 to $200.

Prepaid restructure: Pay As You Go / Starter $500 / Growth $2,000

Oct 2022

Shifted to a prepaid-credit model. Pay As You Go and Starter (pre-paid $500/yr) kept base rates Pre-recorded $0.0125/min, Live streaming $0.0158/min. Growth (pre-paid $2,000/yr) discounted Pre-recorded to $0.0121/min and Live streaming to $0.0125/min. Understanding add-ons +$0.0083/min, extra audio channels +$0.0040/min across plans.

Three-plan model: Try Free / Standard / Premium

Jan 2022

Standard transcription billed per minute: Batch $0.0125/min, Streaming $0.0158/min. Try it Free gave $150 in free credits, no credit card. Add-ons priced per minute — Diarization $0.0016/min, Deep Search $0.0016/min, extra audio channel $0.0008/min, Redaction $0.0008/min. Premium was 'Volume Discounts for Committed Usage' (custom).