AssemblyAI publishes a speech to text rate of $0.15 an hour. A clinical configuration of the same API costs $0.46.

Nothing is hidden to get from the first number to the second. Every line item sits on the same pricing page. They just sit below the rate card, where cost estimates tend not to reach.

Budget about 19 cents for a 15 minute patient encounter. The headline rate suggests under four.

Here is where the rest of it goes.

Key Takeaway: AssemblyAI's published speech to text rate starts at $0.15 an hour, but a clinical configuration does not run at the headline rate. Medical Mode adds $0.15 an hour, streaming speaker diarization adds $0.12, and key-terms prompting adds $0.04, which puts the cheapest ambient clinical setup at $0.46 an hour, or 3.1 times the published figure. Two billing rules push it further: streaming is billed by open-to-close session time rather than audio sent, and multichannel audio is billed per channel. Budget roughly $0.19 for a 15 minute encounter rather than the $0.0375 the headline suggests. None of those rates include turning that transcript into a clinical note.

Why the AssemblyAI Headline Rate Does Not Survive a Clinical Encounter

The $0.15 is real. It is one of the lowest base rates in the category, and AssemblyAI is not doing anything unusual by charging for features separately.

The problem is that the three features a clinical product cannot ship without are all separately charged, and together they cost twice the base rate.

The Published Rate and What It Covers

Pre-recorded audio runs $0.15 an hour on Universal-2 and $0.21 on Universal-3.5 Pro. Streaming runs $0.15 an hour on either Universal-Streaming model, or $0.45 on Universal-3.6 Pro Realtime [1].

That base rate buys a transcript. Not a speaker-attributed transcript, not a medically tuned one, and not one that knows your drug list.

Pre-recorded and streaming are the same $0.15 at the entry tier, which makes the choice between them look free. It is not, and the reason is further down this page.

The free tier needs no card and is metered in hours rather than dollars: up to 185 hours of pre-recorded transcription and up to 333 hours of streaming [1]. Because the allowance is hours of audio, turning on Medical Mode does not shrink it, which makes the free tier a usable way to test a clinical configuration rather than only a base one. Those 333 streaming hours are roughly 1,300 fifteen-minute visits.

The Three Line Items a Clinical Configuration Adds

Start with Universal-Streaming English at $0.15 an hour, the cheapest path to an ambient clinical product.

Medical Mode, $0.15 an hour. Switched on with "domain": "medical-v1", it is the model tuned for clinical speech [1][2]. It doubles the base rate on its own.

Streaming speaker diarization, $0.12 an hour. A note needs to know which words came from the clinician. The async version of the same feature is $0.02, so the live version costs six times more [1].

Keyterms prompting, $0.04 an hour. Your own formulary, passed into the call. Included free on Universal-3.6 Pro Realtime and Universal-Streaming Multilingual, charged on Universal-Streaming English, which is the model an English-only product would otherwise pick [1].

That is $0.46 an hour, against a $0.15 headline. None of the three is optional for clinical audio.

What One Encounter Actually Costs

A 15 minute visit at $0.46 an hour is $0.115. On the premium realtime model it is $0.18. Recorded rather than streamed, it is $0.0925.

Read from the headline alone, you would have written down $0.0375.

The async figure is lower for a reason worth noting early: async diarization is $0.02 against $0.12 for the streaming version [1]. If your product can transcribe after the visit rather than during it, that one decision moves the bill more than any other line on this page.

The closest anyone else gets to this arithmetic is Gladia, whose AssemblyAI pricing analysis ranks sixth for this term and is cited in Google's AI answer [3]. Gladia makes the right general argument, that add-on pricing is hard to forecast, then builds its example from sentiment analysis, summarization and topic detection. No clinical product turns any of those on, and summarization is now deprecated [1]. The words medical, clinical and Medical Mode do not appear on the page.

The Two Billing Rules That Change the Answer

Both of these are stated plainly on AssemblyAI's pricing page. Neither appears in any third-party analysis of it, and both change a clinical forecast more than the add-ons do.

Streaming Bills Session Time, Not Audio

Streaming is billed on the total time your WebSocket connection stays open, not on the audio you send [4]. Idle time counts, and a session you never close auto-closes after three hours and bills for all three [4]. At the clinical rate that is $1.38 for one connection nobody terminated.

Think about what that means in an exam room. If the connection opens when the clinician walks in and closes when they walk out, you are billed for the chart review, the silence while the patient undresses, and the conversation they have with the nurse.

A 15 minute conversation inside a 25 minute session bills as 25 minutes. That turns $0.115 into $0.1917, which is 5.1 times the headline estimate.

Two things to do about it. Open the socket when recording starts rather than when the app loads, and close it on the first silence after the visit ends. Then log session duration and audio duration as separate metrics, because the gap between them is a cost line nobody will find otherwise.

Multichannel Audio Bills Per Channel

Each audio channel is billed separately. AssemblyAI's own example: a one hour stereo file on Universal-3.5 Pro costs 2 x $0.21 = $0.42 [1].

Clinical rooms often record two channels, a lapel mic and a room mic. Telephony is worse, because call recordings routinely put each party on its own channel.

If your pipeline submits stereo without downmixing, your transcription bill is exactly double and nothing in the response says so. Downmix to mono unless you are using the channels for speaker separation, in which case you are paying for diarization twice over.

AssemblyAI Pricing per Minute Across Every Model

AssemblyAI publishes hourly rates. Products get budgeted per encounter. Here is the whole rate card both ways, current as of 4 November 2026 [1].

Pre-Recorded Models and Rates

Pre-recorded is billed per hour of audio submitted, which is the intuitive behaviour. If you are on an older model identifier, check it against the current rate card before you forecast, because the published line-up has changed more than once this year [1].

Streaming Models and Rates

Every rate in this table is billed by connection time rather than audio length, so treat these as a floor [4].

Medical Mode, Diarization and Key-terms

The three a clinical build turns on, with the per-model variation that decides which base model you should pick.

The term limits are the part worth reading twice. A working formulary plus the specialty vocabulary and the referring physicians' names runs past 200 entries quickly, and the streaming models cap at 100. AssemblyAI reports that Medical Mode cuts missed clinical entities against its general model [6], which is the argument for paying for it rather than leaning on keyterms alone.

Two More Line Items Worth Checking Before You Build

Most of AssemblyAI's catalogue is irrelevant to a clinical pipeline, but three entries are not.

PII redaction, $0.05 an hour on audio and $0.08 on text. On streaming the text figure is $0.12, not $0.08 [1]. If you are de-identifying transcripts for analytics or model training, price the streaming number, because that is the mode an ambient product runs in.

Voice Focus, $0.10 an hour. Isolates the primary speaker and suppresses the rest of the room, on streaming models only [1]. No ambient clinical product should price a configuration without deciding on this one, because an exam room is exactly the problem it solves.

Check the catalogue before you copy an older total. Auto Chapters no longer appears on the pricing page at all, and Summarization is priced per hour rather than per model the way the add-on tables are [1]. Any cost comparison built on the older line-up is describing a different product. Gladia's analysis is one of them [3].

What to Budget per Encounter and per Month

A 15 Minute Visit, Priced Five Ways

Budget on the last row. It is the only one that reflects both the features a clinical product needs and the way AssemblyAI actually meters a live connection.

Monthly Cost as Volume Grows

At 1,000 encounters a month the headline math says $37.50 and the realistic figure is $191.67. At 10,000 it is $375 against $1,916.67.

Those are not large numbers, and it would be easy to overstate them. Transcription is rarely the line item that decides a health-tech budget.

The problem is not the amount. It is that a model built on the headline is wrong by a factor of five, and that error compounds into every unit-economics slide, every pricing decision and every gross-margin forecast that sits downstream of it.

It also tends to surface at the worst moment. The base rate is what gets into the spreadsheet during evaluation, when nobody has turned Medical Mode on yet and the prototype streams thirty seconds of test audio. The session-duration behaviour only shows up once real visits are running, which is usually after the pricing page is public.

How AssemblyAI Pricing Compares Once Clinical Features Are On

Where the Headline Rank and the Clinical Rank Diverge

Configured for clinical audio and compared streaming against streaming, per minute: AssemblyAI $0.0077, Deepgram $0.0110 once streaming diarization and keyterm prompting are added [7], Speechmatics $0.0038 on its Oak 1 medical model, which includes diarization [8].

AssemblyAI starts cheapest and finishes at twice the Speechmatics rate. Deepgram starts highest and stays highest. Speechmatics moves least: Oak 1 real time is $0.23 an hour, a sliver above the $0.20 it charges for its general multilingual model and well under the $0.40 for its feature-rich one [8]. The vendor with the lowest headline rate on this chart is not the one with the lowest clinical rate.

Two caveats on that Speechmatics figure. Custom dictionary is not offered on Oak 1, so the configurations are not strictly equivalent, and Speechmatics was serving different rates to different sessions on the day this was written, which is a fair sign the price is still moving. Confirm it against your own account before you budget on it [8].

The two general cloud vendors are not close. Google charges $0.078 a minute for its medical models [9] and Amazon Transcribe Medical works out to $0.075 [10], roughly ten times the specialists.

What the Comparison Still Does Not Price

Every rate on this page buys a transcript.

None of them buys a note, a set of fields validated against your schema, or a write into the EHR. That work is engineering time on your side of the API, it does not appear in any per-minute comparison, and it is usually larger than the difference between any two vendors here.

Put the two in proportion. Moving from the most expensive specialist on this chart to the cheapest saves $0.0072 a minute, which is about $108 a month at 1,000 fifteen-minute encounters. One engineer spending a quarter of their time maintaining the transcript-to-note layer costs more than that in a week.

That is the comparison worth running, and no pricing page on either side of it will give you the numbers.

How Sully Prices the Same Workload

Sully does not publish a per-minute rate, so this article cannot put Sully on the chart. In a piece about reading someone else's price list, pretending otherwise would be the exact move it criticises.

What can be compared is the shape of the bill.

Word boosting is part of streaming transcription rather than a line item on top of it [11]. Note generation takes a transcript and your own template, and Text to JSON takes your JSON Schema and returns validated fields, so the work between a transcript and an EHR-ready record is an endpoint rather than something you build and maintain [11]. Zero Data Retention is the published default, with nothing stored after processing and no training on customer data [11], under ISO 27001, HIPAA, HITRUST and SOC 2 Type II [12].

For a number against your own volume, get an API key at dashboard.api.sully.ai and run your own encounters through it. There is no demo to sit through [11].

FAQ

How much does AssemblyAI cost per minute?

Pre-recorded audio is $0.0025 a minute on Universal-2 and $0.0035 on Universal-3.5 Pro. Streaming is $0.0025 a minute on either Universal-Streaming model and $0.0075 on Universal-3.6 Pro Realtime [1]. AssemblyAI publishes hourly rates, so those are converted. For clinical audio the figure to budget on is higher, because Medical Mode, speaker diarization and keyterms prompting are billed separately and stack on the base. The cheapest ambient clinical configuration works out to about $0.0077 a minute.

What are the limitations of AssemblyAI?

For a clinical build, three are worth knowing before you commit. Streaming is billed by session duration rather than audio sent, so idle connection time is billable [1]. Multichannel audio is billed per channel, so a stereo recording costs double unless you downmix [1]. And keyterms prompting caps at 200 terms on Universal-2 and 100 on the streaming models, which is below a working formulary [1]. None of these is hidden. All three are on the pricing page, and most cost estimates skip them.

Does AssemblyAI have a medical model, and what does it cost?

Yes. Medical Mode is enabled with "domain": "medical-v1", and it costs an extra $0.15 an hour on every model that supports it [1][2]. It covers English, Spanish, German and French, except on Universal-Streaming English where it is English only [1][5]. Because it is an add-on rather than a separate model, it stacks on whatever base rate you are already paying, which doubles the cost of the $0.15 an hour tiers.

Can AssemblyAI transcribe phone calls?

Yes, both recorded and live. For clinical telephony the thing to check first is the channel count. Call recordings frequently put each party on a separate channel, and AssemblyAI bills per channel, so a one hour two-party call is billed as two hours [1]. That is worth paying for if you want the channels for speaker separation, and worth avoiding if you do not.

Is AssemblyAI HIPAA compliant?

AssemblyAI lists a Business Associate Agreement under its enterprise plan, so the compliance path exists as a contract conversation rather than a self-serve toggle [3][13]. As with any vendor, the more useful question is what the service retains by default before anything is negotiated, and that answer belongs in writing. Sully's APIs run Zero Data Retention, with nothing stored after processing and no training on customer data [11], under ISO 27001, HIPAA, HITRUST and SOC 2 Type II [12].

Sources

[1] AssemblyAI— Pricing [2] AssemblyAI — Medical Mode, streaming [3] Gladia — AssemblyAI pricing: Is it worth it or consider Gladia?[4] AssemblyAI— Streaming Speech-to-Text quickstart [5] AssemblyAI — Medical Mode, pre-recorded [6] AssemblyAI — Introducing Medical Mode[7] Deepgram — Pricing [8] Speechmatics— Pricing [9] Google Cloud — Speech-to-Text pricing [10] Amazon Web Services — Amazon Transcribe pricing [11] Sully.ai — Sully APIs [12] Sully.ai — Trust Center [13] ColdIQ — Assembly AI Review