Deepgram

What Deepgram is

Deepgram is speech infrastructure, not an app. There is no editor, no dashboard to hand to a marketing team, no export button. You send audio to an API and get text back, you send text and get audio back, or you open a websocket and let Deepgram run a whole voice conversation. Everything is billed per minute of audio or per thousand characters of text, and the rates are printed on a public page for both self-serve plans rather than hidden behind a sales call. In a category where "contact us" is the norm, that alone is worth something.

The catalogue splits four ways. Speech-to-text runs on Nova-3, the general-purpose model, which the pricing page says covers 45+ languages and the product page says covers 50+ (the kind of drift you get when two teams own two pages), and on Flux, a conversational model built for voice agents with turn detection and interruption handling built in. Flux Multilingual shipped on 29 April 2026 covering ten languages with mid-conversation switching. Deepgram also hosts Whisper Large at a flat rate, and keeps Nova-2, Enhanced and Base alive at their old prices for existing deployments. Text-to-speech is Aura-1, Aura-2 and Flux TTS. Audio Intelligence adds summarization, topic detection, sentiment analysis and intent recognition, billed per thousand tokens rather than per minute. The Voice Agent API packages all of it: one websocket that listens, thinks and speaks, with tiers for bringing your own LLM, your own TTS, or both.

The company is not a startup gamble at this point. Deepgram closed a $130 million Series C at a $1.3 billion valuation on 13 January 2026, led by AVP, with In-Q-Tel, BlackRock-managed funds, Twilio, ServiceNow Ventures, SAP and Citi Ventures among the investors, and folded in OfOne, a drive-through voice startup, in the same announcement. Its April 2026 press release claims more than 200,000 developers and 1,400 organizations (the January release said 1,300, so the number is moving and self-reported). It is SOC 2 Type 2 certified, signs a HIPAA BAA for Enterprise customers, is PCI and GDPR compliant, and offers an EU endpoint at api.eu.deepgram.com plus self-hosted containers for customers who cannot send audio outside their own VPC.

What you are buying is a well-capitalised, boring pipe. The complications are all in how it is priced, and that is where most of this review lives.

Pricing

Deepgram publishes two full rate cards, Pay As You Go and Growth, and every line below carries both. The third tier, Enterprise, is contact-sales with nothing published. Pay As You Go starts with a $200 credit, no credit card required, no minimums, no expiration, then bills against usage. Growth is $4K+ a year, described by Deepgram as up to 20% savings on pre-paid credits redeemed against actual usage. Per-second billing applies throughout: Deepgram's own FAQ says a 14-second file bills as 14 seconds.

Speech-to-text, streaming

The streaming table is headed "Limited-time promotional rates on streaming" and no end date is published for it. Legacy models, per the pricing FAQ: Nova-2 streaming $0.35 an hour, Enhanced $0.99 an hour, Base $0.87 an hour, with Growth about 12.5% lower.

Speech-to-text, pre-recorded

No promotional markdown on this table. This is the cheapest genuinely stable pricing on the page.

Speech-to-text add-ons

Text-to-speech

During the free window, Flux TTS allows up to 45 concurrent streaming connections globally and 5 in EU/AU.

Voice Agent API

Every cell in this table is footnoted "calculated based on websocket connection time".

Audio Intelligence

All figures in US dollars, read from deepgram.com/pricing in September 2026. The rate tables and plan comparison render fully with nothing behind a login; the FAQ answers sit in collapsed accordions, and the billing mechanics quoted on this page were read from that FAQ text in the page source.

How to read the promotional columns

A large share of the headline rates are promotional, and the promos are not all the same promo. Only two dates appear on the page, 12 September 2026 and 13 September 2026, and both attach to Flux TTS ("free until 9/12") and the Voice Agent tiers ("through 9/12"). The streaming speech-to-text table carries the line "Limited-time promotional rates on streaming" and prints a current price beside a regular price, but publishes no date for when current becomes regular. The increase is scheduled and disclosed, but undated.

The practical takeaway: model the regular column, not the current one. Deepgram's own FAQ quotes Nova-3 at $0.29 an hour for monolingual streaming and $0.35 for multilingual, which is the current promotional rate times 60. At the regular rate the same hours cost $0.46 and $0.55. If your business case only works at the current price, you do not have a business case yet.

What we verified, and what we did not

Reviews of API vendors often imply that a score covers accuracy. This one does not, so here is the boundary in plain terms.

Verified: every price and plan term above, read from deepgram.com/pricing in September 2026; the Series C amount, valuation, date and investor list from Deepgram's own press release; the Flux Multilingual launch date and language count from its press release; the 50+ language and sub-300 millisecond claims from the speech-to-text product page; the $4.50 an hour Voice Agent figure from its product page; the EU endpoint hostname and compliance list from the pricing page and Deepgram's trust documentation; and the Enterprise-only, NVIDIA GPU requirement for self-hosting from the pricing FAQ and the developer docs.

Not verified: accuracy. We ran no transcription of our own, so every word error rate, latency and keyword recall figure here is Deepgram's claim. Real-world uptime: no numeric SLA is published for either self-serve plan. Whether idle websocket time bills on the Voice Agent API: we did not open a socket and measure it, so the conflict described below is reported as a documentation discrepancy, not a measured overcharge. Enterprise pricing, self-hosted licence cost and whether anything is negotiable below the $4K Growth entry point: all behind sales. And revenue: a figure above $100 million ARR appeared in August 2026 trade coverage, sourced to Deepgram, and is not confirmed by any filing, so it is not in this review.

Plan comparison: what $4K a year actually changes

The Growth card headline is "Save up to 20%". Here is what the comparison table on the same page says the money buys, and what it does not.

Two things stand out. First, paying $4K a year does not buy you a support channel. The plan comparison lists Community and Discord for both self-serve tiers, with the same Standard Uptime SLA. The FAQ mentions priority support for Growth, but the channel list is identical, and a real support relationship starts at Enterprise. Second, exceeding your Growth credits does not drop you to Pay As You Go rates. Per the FAQ, overages bill at the Growth rate plus a 10% premium plus tax, calculated every Monday for the previous week; a card must stay on file, and if you remove the card and exhaust credits the API returns a 402 until you top up. The escape hatch is what Deepgram calls an early renew: buy a fresh Growth plan through the console to reset credits at the standard rate.

Where Deepgram is the strongest pick

Deepgram is at its strongest for a team that already has engineers and wants a rate card it can read without a sales call. High-volume pre-recorded transcription is the clearest case: Nova-3 monolingual at $0.0043 a minute is $0.26 an hour on Pay As You Go and $0.22 an hour on Growth, with diarization and smart formatting included, and no promotional asterisk on any of it. Per-second billing makes that number real on short clips, where a vendor rounding to 15 seconds or a full minute inflates a bill of IVR fragments or voicemail by a wide margin.

The second case is the regulated buyer. SOC 2 Type 2, a HIPAA BAA for Enterprise customers, PCI, a named EU endpoint and self-hosted containers are a compliance story you can hand to a security reviewer as-is, and the pricing page says so in plain language rather than in a vague GDPR claim.

The third is a voice agent with an LLM you already pay for. Custom BYO LLM plus TTS at a flat $0.050 a minute, or $0.041 on Growth, carries no promotional cliff and is a third off the Standard tier's regular rate.

Best for

Engineering teams building transcription or voice-agent features into their own product, who want published per-minute rates, per-second billing and a compliance story they can hand to a security reviewer. Particularly strong for high-volume pre-recorded transcription, where $0.0043 a minute is cheap and stable, and for regulated buyers who need an EU endpoint, a HIPAA BAA or self-hosted containers. Not for anyone who wants a finished product: there is no editor, no meeting bot and no interface, and a team that wants transcripts rather than an API should be looking at Otter.ai, HappyScribe or Descript instead.

Key features

• Nova-3 speech-to-text, 45+ languages per the pricing page and 50+ per the product page, with speaker diarization, smart formatting, keyterm prompting and automatic language detection
• Flux, a conversational speech recognition model with built-in turn detection and interruption handling; Flux Multilingual covers ten languages with mid-conversation switching, generally available since 29 April 2026
• Voice Agent API on a single websocket, bundling listening, LLM orchestration and speaking, with BYO LLM and BYO TTS tiers that cut the rate
• Text-to-speech across Aura-1, Aura-2 and Flux TTS, billed per 1,000 characters
• Audio Intelligence: summarization, topic detection, sentiment analysis and intent recognition, billed per 1,000 tokens rather than per minute
• Hosted Whisper Large at $0.0048 a minute on both self-serve plans, for teams that want Whisper output without running GPUs
• True per-second billing with no rounding, stated in Deepgram's own pricing FAQ
• PII redaction, entity detection and keyterm prompting as priced add-ons; smart formatting and pre-recorded diarization included at no charge
• SOC 2 Type 2, HIPAA BAA for Enterprise customers, PCI, CCPA and GDPR, with a dedicated EU endpoint at api.eu.deepgram.com
• Self-hosted containers for VPC or on-premise deployment on the Enterprise plan, requiring NVIDIA GPUs
• Legacy Nova-2, Enhanced and Base models kept at unchanged rates for existing deployments
• $200 starting credit with no credit card, no minimums and no expiration

Best-fit use cases

• High-volume pre-recorded transcription of call recordings, podcasts or media archives at $0.0043 a minute on Nova-3 monolingual, or $0.0036 on Growth
• A real-time voice agent for a contact centre or phone line on the Voice Agent API, with an existing LLM brought in on the Custom BYO tier to cut the per-minute rate
• Live captioning or real-time transcription inside a meeting, streaming or telephony product
• Call-centre analytics: transcription plus Audio Intelligence for sentiment, intent and topic detection, priced per token instead of per minute
• Regulated transcription in healthcare, legal or finance where a HIPAA BAA, an EU endpoint or self-hosted containers are hard requirements
• Replacing self-managed Whisper GPU infrastructure with hosted Whisper Large at $0.0048 a minute
• Redacting PII from recorded audio at $0.0020 a minute before it reaches downstream storage

Pros

• Publishes complete per-minute rates for every model on both self-serve plans, Voice Agent included, which most of this category will not do without a sales call
• Per-second billing with no rounding, a genuine cash difference against vendors that round to the nearest 15 seconds or full minute
• Pre-recorded Nova-3 monolingual at $0.0043 a minute ($0.0036 on Growth) is among the cheapest published transcription rates anywhere and carries no promotional asterisk
• The $200 starting credit needs no credit card and does not expire, so evaluation costs nothing and is not on a clock
• Voice Agent BYO tiers are real discounts rather than theatre: Custom BYO LLM plus TTS at $0.050 a minute is a third off the Standard regular rate, with no promo cliff
• Compliance posture is enterprise-grade and specific: SOC 2 Type 2, HIPAA BAA, PCI, and a named EU endpoint rather than a vague GDPR claim
• Self-hosted deployment exists for regulated buyers who cannot send audio to a third party
• Old models stay available at old prices, so an existing Nova-2 deployment is not force-migrated
• Well funded and widely deployed: a $130M Series C at a $1.3B valuation in January 2026, with strategic money from Twilio, ServiceNow, SAP and Citi

Cons

• Much of the headline pricing is promotional, and the speech-to-text streaming promo has no published end date, so the largest line in most budgets is on an undated countdown
• Regular streaming rates are 59 to 60% above the current ones for Nova-3 on Pay As You Go, and 55 to 56% above on Growth, which is a large increase to have sitting in a column labelled regular price
• Voice Agent tiers are footnoted as billed on websocket connection time, while the FAQ on the same page describes billing on conversation duration; those are different meters when a socket sits open through silence or hold
• Support is Community and Discord on both self-serve plans; paying $4K a year for Growth does not add an email channel, and a real support relationship starts at Enterprise
• Growth overages bill at the Growth rate plus a 10% premium rather than dropping to Pay As You Go, and are invoiced weekly
• Auto-reload is on by default at $100 when the balance drops to $10, so a runaway job spends real money without a prompt, and Deepgram's FAQ puts the responsibility on you
• Unused purchased credits are refundable only within 30 days; anything older is gone, and Deepgram states you pay for faulty or looping files your own system uploads
• Multichannel audio bills per channel, so a stereo recording costs double what the file duration suggests
• Nothing here is usable without engineering; there is no interface for a non-developer, no meeting bot and no editor
• Language-count claims disagree between the pricing page (45+) and the product page (50+)
• After the promo, Nova-3 monolingual streaming at $0.0077 a minute ($0.46 an hour) costs more than legacy Nova-2 at $0.35 an hour, so the upgrade path is a 32% price rise

What to watch out for

Five things, in order of how much money they cost you.

Model the regular column, not the current one. The streaming speech-to-text table shows a current price and a regular price side by side under the line "Limited-time promotional rates on streaming", and gives no date for the switch. Nova-3 monolingual streaming is $0.0048 a minute today and $0.0077 regular, a 60% increase; multilingual goes from $0.0058 to $0.0092. Growth is not exempt: $0.0042 becomes $0.0065. Voice Agent and Flux TTS at least name their date: promotional rates run through 12 September 2026 and standard pricing begins 13 September, ten days after this review was written.

The Voice Agent meter is the websocket, not the conversation. Every Voice Agent cell in the rate table carries the footnote "calculated based on websocket connection time". The FAQ lower on the same page says the API "is billed based on the duration of the conversation". Those agree only if you close the socket the moment nobody is talking. If your agent holds a connection through hold music, an IVR transfer, or a caller who has stopped speaking but not hung up, you are paying $0.075 a minute for silence once the promo ends. Instrument connection lifetime before you forecast anything.

The $200 credit does not buy 700 hours at the price you will pay. Deepgram's FAQ says the credit is approximately 43,000 minutes, over 700 hours, of transcription on Nova. That works out to about $0.0047 a minute, which sits between the pre-recorded rate and the promotional streaming rate. At the regular streaming rate of $0.0077 the same $200 buys about 25,970 minutes, or 433 hours; at Voice Agent Standard's regular $0.075 it buys about 44 hours. Not a false claim, but it is measured against the cheapest configuration and it will not describe a production streaming workload.

Auto-reload is on by default, and support is Discord at every self-serve tier. New accounts reload $100 automatically when the balance falls to $10, and the FAQ says plainly that managing this is the customer's responsibility. Combine that with a bug that loops audio, the explicit disclaimer that you pay for faulty files your own system uploads, a 30-day window on refunds for unused purchased credits, and community-only support on both Pay As You Go and Growth, and a bad weekend is expensive and slow to resolve. Turn auto-reload off during development and set your own alerting.

Re-check the Voice Agent wiring on 13 September. During the promo, Voice Agent Standard at $0.056 a minute is cheaper than Standard BYO TTS at $0.065, because Flux TTS is free until 12 September. After that date it inverts, $0.075 against $0.065. Whichever way you wire it now, redo the arithmetic when the standard rates land.

Is Deepgram worth it?

Yes, for the buyer it is built for. An engineering team that needs transcription or a voice pipeline inside its own product gets published rates on both self-serve plans, per-second billing, a compliance list a security reviewer can accept, and a $200 credit that lets them find all of this out for free and without a clock. Pre-recorded Nova-3 at $0.0043 a minute is close to the cheapest stable price in the category, and hosted Whisper at $0.0048 removes a GPU fleet from someone's to-do list.

It scores 4.4 rather than higher for two reasons, and neither is output quality. The first is the undated streaming promotion: a 59 to 60% increase sitting in a column labelled regular price is disclosed but not scheduled, and that is the biggest single line in most transcription budgets. The second is support. Community and Discord are the only channels on both Pay As You Go and Growth, so a customer paying $4K a year with auto-reload on, a 10% overage premium and a 30-day refund window carries all the operational risk with nobody to email. Dating the promo or giving Growth a support channel would each move the score.

Worth saying plainly: we did not measure accuracy. Deepgram's sub-300 millisecond transcript and sub-400 millisecond end-of-turn claims are the vendor's, and the score weights what we could verify from the published record. If a team runs its own word error rate comparison on its own audio, the score can move in either direction.

How we arrived at 4.4

The score is weighted across the eight dimensions in our methodology. Each is graded on what we could verify from Deepgram's own published material; nothing here reflects a benchmark we ran.

4.4 is a category leader with an unusually honest pricing page that is nonetheless writing cheques its regular column will cash. Above 4.5 would require either dating the streaming promo or giving the self-serve plans a support channel.

FAQ

Pricing, plan terms and feature claims verified September 2026 from deepgram.com/pricing (including its FAQ text), deepgram.com/product/speech-to-text, deepgram.com/product/voice-agent-api, Deepgram's Series C and Flux Multilingual press releases, and developers.deepgram.com. We ran no transcription tests. ToolChase is not an affiliate of Deepgram and received no payment, credits or access for this review. Scores, rankings and list positions are never sold. See our methodology for how we score tools.