Quick answer: For a live phone agent, Deepgram Flux English streams at $0.0065 a minute (promotional pricing, which the pricing page’s data ends on 13 February 2027; $0.0077 list) and AssemblyAI-STT-3.7-Realtime-Agent at $0.45 an hour ($0.0075 a minute). At 10,000 call minutes that is $65.00 against $75.00, but $78.00 against $75.00 once Deepgram’s $0.0013 keyterm add-on is on, because AssemblyAI includes keyterms. Flux takes turns in 10 languages, English included; AssemblyAI in 32.
“Deepgram vs AssemblyAI” is usually asked by a team wiring speech-to-text (STT) into a phone agent: the caller’s audio arrives from a carrier such as Twilio as 8 kHz mu-law, the transcript has to come back fast enough for a large language model to answer, and the engine has to say when the caller has finished. This page compares only that job: streaming STT on live calls. It leaves out batch transcription, meeting notes and call analytics. On Deepgram it covers Flux (the voice-agent model with built-in end-of-turn detection) and Nova-3 (its general model, which streams but has no turn detection). On AssemblyAI it covers AssemblyAI-STT-3.7-Realtime-Agent, released on 18 November 2026, its notetaker sibling AssemblyAI-STT-3.7-Realtime, and Universal-3.6 Pro Realtime, which AssemblyAI says is still fully supported. Every figure was read on deepgram.com, developers.deepgram.com, assemblyai.com and assemblyai.com/docs on 22 November 2026, in US dollars.
Where Zian fits: an AI agents platform, not a speech-to-text API
If what you need is calls answered and appointments booked rather than a transcription pipeline to build and tune, Zian AI is the platform to look at first, because it takes the speech-to-text decision off your desk. LeadsNow.ai is the done-for-you route. Zian AI is a standalone AI sales and reception agents platform, not a phone system (carrier / CPaaS / contact centre) and not an add-on to one.
Zian answers inbound calls 24/7 as an AI receptionist that qualifies, books, transfers and takes messages, runs outbound calls, and follows up by SMS, email, WhatsApp, Instagram DMs and Facebook Messenger, with answering-machine and voicemail handling on calls. It works in 30+ languages, supports voice cloning, connects to HubSpot, Salesforce, HighLevel and Zapier, and offers private model deployment on customer infrastructure. SmartReach AI™ orchestrates the message, channel and timing of follow-up, and PrecisionPitch AI™ runs continuous split-testing of scripts and approaches. Zian has set 64,521+ qualified sales appointments. LeadsNow.ai runs on Zian as a managed, pay-per-result service that answers leads in under 60 seconds, 24/7, and has set 50,769+ appointments. Zian is in a partnership-application beta with no published prices; each of its three listed plans, including Intel + Widgets (Free Forever), starts with a partnership application.
Apply For Partnership to see whether Zian’s inbound and outbound agents fit your call volume.
What to compare when the audio is a live phone call
Batch accuracy tables tell you little about a phone agent. Six questions decide it, and this page answers each one from the two vendors’ own pages:
- What a streaming minute costs, on what basis it is metered, and which features cost extra.
- Whether the engine decides when the caller has finished, and what you can tune when it gets that wrong.
- Which languages work in a live stream, as opposed to the longer batch list on the home page.
- Whether 8 kHz mu-law from the carrier goes in as it is, or has to be converted first.
- How many calls can run at once before you hit a limit.
- Where the audio is processed, what is kept, and whether you can run the model yourself.
Deepgram vs AssemblyAI streaming STT, side by side
Each cell comes from the vendor’s own pricing page or documentation, read on 22 November 2026. Prices are pay-as-you-go list rates.
| Question | Deepgram (Flux and Nova-3) | AssemblyAI (STT-3.7-Realtime-Agent and siblings) |
|---|---|---|
| Streaming price, pay as you go | Flux English $0.0065 a minute (was $0.0077); Flux Multilingual $0.0078; Nova-3 Monolingual $0.0048 (was $0.0077); Nova-3 Multilingual $0.0058 (was $0.0092). The table is badged “Promo pricing” (the page data ends the promotion on 13 February 2027); Flux Multilingual shows no “was” price | STT-3.7-Realtime-Agent $0.45 an hour; STT-3.7-Realtime $0.30 an hour; Universal-3.6 Pro Realtime $0.45 an hour (models page) |
| What is metered | “Billed per second, with no block-based rounding for normal audio requests” | The time the WebSocket stays open, “not on the amount of audio you send”; an unterminated session closes after 3 hours and bills all 3 |
| Keyterm prompting | Add-on, $0.0013 a minute; up to 100 terms and 500 tokens; Nova-3 and Flux | Included on both STT-3.7 streaming models; up to 100 keyterms, each 50 characters or less (pricing page: maximum 6 words per phrase) |
| Other streaming add-ons | Redaction $0.0020, Entity Detection $0.0017, Speaker Diarization $0.0020 a minute; Smart Formatting included | Contextual prompting $0.05, Voice Focus $0.10 (Agent model only), Speaker Diarization $0.12, PII text redaction $0.12, Medical Mode $0.15 an hour |
| Turn detection from the engine | Flux: model confidence (eot_threshold 0.5 to 1.0) with a 500 to 60,000 ms silence backstop. Nova-3: none; silence endpointing only | All three models: transcript-based end-of-turn checks with entity holds; manual mode for your own detector on the Agent model only |
| Languages with turn detection in one stream | English on flux-general-en; 10, English included, on flux-general-multi | 32 on every streaming model |
| Languages that stream at all | Nova-3: 99 single languages, 55 in one multilingual stream | The same 32 |
| 8 kHz mu-law in | Yes: mulaw and alaw raw, sample_rate set by you | Yes: pcm_mulaw; sample_rate any integer from 8000 to 96000 |
| Concurrent streams, pay as you go | Up to 150 per project (Flux and Nova-3, every region) | No cap on open sessions; 200 new sessions a minute in the docs, 100 in the pricing FAQ, scaling up automatically |
| Processing regions | Global, EU, Australia and India endpoints | US and EU only, or edge routing across both |
| Self-hosting | Enterprise; containers on your own VPC or hardware with NVIDIA GPUs; Flux included | Self-hosted streaming with “an upfront commercial commitment of $50,000” |
The quotable version: Deepgram sells the cheaper English minute while its promotion runs, and the wider plain-transcription language list, while AssemblyAI includes keyterms in its rate and takes turns in more than three times as many languages.
What does 10,000 call minutes cost on Deepgram vs AssemblyAI?
This is the worked bill. Every line is a published rate times a quantity you can change.
Assumptions, labelled. A1: 10,000 billed minutes a month, one stream per call carrying the caller’s side only. A2: on AssemblyAI, each session opens when the call connects and is terminated at hang-up, so session minutes equal call minutes. A3: on Deepgram, the carrier streams audio for the whole call, so billed audio seconds equal call seconds. A4: pay-as-you-go rates, no volume discount, no Growth or Enterprise plan. A5: no diarization, redaction or medical mode. A6: carrier and LLM costs are left out; they are the same whichever STT you pick. A7: Deepgram’s keyterm add-on is billed on Flux as on Nova-3; its pricing page lists the add-on without exempting any model, so check your first invoice.
| Line, 10,000 minutes | A. Deepgram Flux English, promo | B. Flux English at list | C. Flux Multilingual | D. Nova-3 Monolingual, promo | E. AssemblyAI STT-3.7-Realtime-Agent |
|---|---|---|---|---|---|
| Rate | $0.0065 / min | $0.0077 / min | $0.0078 / min | $0.0048 / min | $0.45 / hr = $0.0075 / min |
| Transcription | $65.00 | $77.00 | $78.00 | $48.00 | 166.67 hr × $0.45 = $75.00 |
| Keyterm prompting | $13.00 | $13.00 | $13.00 | $13.00 | Included |
| Monthly total without keyterms | $65.00 | $77.00 | $78.00 | $48.00 | $75.00 |
| Monthly total with keyterms | $78.00 | $90.00 | $91.00 | $61.00 | $75.00 |
| Effective rate with keyterms | $0.0078 / min | $0.0090 / min | $0.0091 / min | $0.0061 / min | $0.0075 / min |
| Engine ends the turn? | Yes | Yes | Yes, 10 languages | No: you build turn logic | Yes, 32 languages |
How we checked it. Deepgram: 10,000 × rate, then 10,000 × $0.0013 for keyterms (route A: $65.00 + $13.00 = $78.00). AssemblyAI: 10,000 ÷ 60 = 166.67 hours × $0.45 = $75.00, which equals 10,000 × $0.0075. Every column was recomputed both ways. Optional AssemblyAI extras on the same 166.67 hours: contextual prompting adds $8.33 and Voice Focus $16.67, which would take column E to $100.00. AssemblyAI’s other realtime model, STT-3.7-Realtime at $0.30 an hour, would bill $50.00, but AssemblyAI says it “detects turns to segment the transcript rather than to drive voice-agent turn-taking”, so it is not a like-for-like agent option.
What the bill says. At 10,000 minutes the five columns run from $48.00 to $91.00 a month, so for most phone agents the STT line is small next to the carrier and the language model. The ordering, though, flips on one add-on and one date. Deepgram’s pricing page marks its speech-to-text rates as promotional, and the page’s own data sets the promotion’s end at 13 February 2027; it does not print the rate that applies after that, so column B uses the “was” price it shows. Deepgram Growth would cut Flux English to $0.0057, but Growth “starts at $4,000 a year in pre-paid annual credits”, which at that rate is about 58,480 minutes a month before the credit is used up; at 10,000 minutes, pay as you go is the plan that fits.
The Keyterm Swing Rule: when Deepgram or AssemblyAI is cheaper
The Keyterm Swing Rule: on English phone calls, Deepgram Flux is the cheaper stream only while you send no keyterms and the promotional $0.0065 holds. Add Deepgram’s $0.0013 keyterm add-on and Flux costs $0.0078 a minute against AssemblyAI’s $0.0075 with keyterms included; at Deepgram’s $0.0077 list rate, Flux costs more with or without keyterms.
The rule is per minute, so it holds at 2,000 minutes or 200,000. What moves it:
- Keyterms are not optional for most booking agents. Caller surnames, street names, product names and drug names are what an agent most often mishears, and both vendors pitch keyterms at exactly that.
- Billing basis. AssemblyAI meters the open WebSocket, idle time included. A session your code forgets to terminate bills 3 hours, or $1.35 on the Agent model. Deepgram meters audio processed, “so a 14-second file bills as 14 seconds”.
- AssemblyAI extras. Contextual prompting ($0.05 an hour) or Voice Focus ($0.10 an hour) each put AssemblyAI back above Flux with keyterms.
- Nova-3 is cheaper again ($0.0048 promotional), but Deepgram describes it as having “no turn detection”, so you pay for that saving in engineering.
If you are still choosing an engine on endpointing rather than price, our comparison of turn detection across Deepgram Flux, ElevenLabs Scribe and Speechmatics runs the same tests on three other engines.
Turn detection: how each engine decides the caller has finished
Deepgram Flux scores every utterance with an end-of-turn confidence and fires EndOfTurn when it crosses eot_threshold, range 0.5 to 1.0, default 0.7. An optional eager_eot_threshold (0.3 to 0.9) sends an early EagerEndOfTurn so your LLM can start drafting, and Deepgram warns that eager mode “can increase LLM API calls by 50-70%”. A silence backstop, eot_timeout_ms, defaults to 5,000 ms and can be set from 500 to 60,000. A Configure message changes all three thresholds and the keyterms mid-call without reconnecting, and ForceEndTurn lets your own detector end a turn. When Flux cuts a caller off, the EndOfTurn event records whether the model, a timeout or your code ended it; our guide to diagnosing Flux cutting callers off with the trigger field walks through it. Nova-3 has no turn model: it finalises on silence, with endpointing on by default at 10 ms and configurable.
AssemblyAI says plainly that “Streaming does not endpoint on silence alone.” After a pause of min_turn_silence, the model reads the transcript so far; a complete sentence ends the turn, and a turn that ends in something still being read out, such as part of a phone number, an email address or a date, is held open until max_turn_silence. The presets set both: min_latency 128 and 640 ms, balanced (the default) 128 and 1,280 ms, max_accuracy 512 and 2,560 ms. AssemblyAI calls this entity-aware endpointing and reports, on its own Universal-3.6 Pro benchmark, phone-number entity errors falling from 7.0% to 2.2% with the median end-of-turn latency unchanged; because roughly one turn in ten is held, the 90th-percentile end-of-turn latency rises from 0.83 to 1.64 seconds. UpdateConfiguration changes the thresholds, keyterms and language list mid-stream. For your own detector there is ForceEndpoint, and a manual mode that switches the built-in detection off, but “Manual mode is supported only by assemblyai-stt-3-7-realtime-agent”.
The practical difference is the shape of the ceiling. Take a caller reading a card number in four groups with 2.4-second gaps. On Flux, each gap sits inside the default 5,000 ms backstop, so the turn survives as long as the model’s confidence stays below 0.7. On AssemblyAI’s balanced preset, each gap is longer than the 1,280 ms forced end, and AssemblyAI says even a held entity ends at max_turn_silence. Its own advice for exactly this case, “while reading a card number off a card”, is to raise max_turn_silence mid-stream, then re-apply the mode. Deepgram gives the equivalent advice for Flux: when collecting “a password, OTP, or account number”, raise eot_timeout_ms and eot_threshold for that segment with a Configure message, then lower them again. We have not measured either engine on such a call; this is what the published parameters allow.
Which languages work in streaming, not just batch
Home-page language counts mostly describe batch. For a live agent, three lists matter, and they are different sizes.
| What you need in a live call | Deepgram | AssemblyAI |
|---|---|---|
| Engine-side turn detection | Flux: English (flux-general-en, “English (all accents)”), or 10 languages on flux-general-multi: English, Spanish, French, German, Hindi, Russian, Portuguese, Japanese, Italian and Dutch | 32 languages on every streaming model, including all 10 of Flux’s plus Arabic, Cantonese, Mandarin, Korean, Hebrew, Persian, Turkish, Urdu, Vietnamese, Marathi and the Nordic languages |
| Code-switching inside one stream | Flux Multilingual across its 10; Nova-3 language=multi across 55, streaming and batch | All 32, on by default; language_codes steers it |
| Plain streaming transcription, one language | Nova-3: 99 languages, listed as 100 rows because English for Australia and New Zealand has its own row; the 36 added on 16 November 2026 cover “batch and streaming workloads” | No separate list: the same 32 |
Of AssemblyAI’s 32 streaming languages, only Norwegian Nynorsk and Xhosa are missing from Nova-3’s single-language list, and Xhosa is in Nova-3’s multilingual stream. So the honest split is this: AssemblyAI takes turns in more languages; Deepgram transcribes more languages, but past Flux’s 10 it leaves turn-taking to you. Deepgram’s pricing page still says “Nova models support 60+ languages”, behind its own 99-language model list. AssemblyAI’s pre-recorded page lists Australian English among its dialects; Flux’s English model covers “all accents”, and Nova-3 takes an en-AU code. For an Australian line, our page on pairing STT and TTS for Twilio in Australian English covers the voice side too.
8 kHz phone audio: does either one need resampling?
Neither does. Twilio Media Streams deliver 8 kHz mu-law, and both engines accept it raw. Deepgram’s Twilio guide opens Nova-3 with encoding="mulaw" and sample_rate=8000 and says this “matches Twilio’s Media Streams format exactly, so the caller’s bytes flow to Deepgram with no resampling”; its encoding page lists mulaw and alaw for Flux too. AssemblyAI’s WebSocket reference lists pcm_mulaw as an encoding and says sample_rate “Accepts any integer from 8000 to 96000”. AssemblyAI’s voice-agent guide, written around its earlier Universal-3.5 Pro Realtime model, adds a telephony tip: “For telephony environments with low-quality audio (such as 8 kHz mulaw), you can prompt the model to tag genuinely unclear segments as [unclear]”, which on its pricing page is the $0.05-an-hour prompting add-on. Its examples default to 16,000 Hz, so set 8000 explicitly.
Concurrency: how many calls can run at once
Deepgram’s rate-limits page allows up to 150 concurrent streaming requests per project on pay as you go, in North America, Europe, Australia and India alike, and 225 in North America on Growth. Two limbs matter: “Secondary projects created on a self-serve account are limited to a single concurrent stream by design”, and streaming diarization is capped lower, at 50 in North America and 25 elsewhere. AssemblyAI says “there is no hard cap on the total number of open streaming sessions”; what it limits is new sessions a minute, which scales up by 10% whenever you use 80% of it. Its own pages disagree on the starting point: the rate-limits page and voice-agent guide say 200 new sessions a minute on pay as you go, while the pricing FAQ says “your starting limit is 100 sessions per minute” and scales at 70%. Plan on 100. Either way, a 10,000-minute month (about 3,300 three-minute calls) is nowhere near either ceiling.
Rather not assemble and tune a voice pipeline at all? Zian runs AI receptionist, outbound calling and messaging agents as one platform. Apply For Partnership
Data residency, retention controls and self-hosting
We do not summarise either vendor’s contract terms here; read the clauses yourself. What the product pages show:
- Where audio is processed. Deepgram runs global, EU, Australian and Indian endpoints, and says of the Australian one that requests “are never routed outside Australia: if the region is unavailable, requests fail rather than fall back to another region.” AssemblyAI says “All streaming audio and transcription data is processed in AWS regions located in the United States and the European Union”, with US and EU data-zone endpoints. For an Australian buyer who needs in-country processing, that rules out AssemblyAI’s cloud API; its self-hosted streaming (the $50,000 commitment below) is the only AssemblyAI route.
- Training and retention switches. Deepgram: a
mip_opt_out=trueparameter per request, after which “Data from opted-out requests is retained only for the duration necessary to process the request.” AssemblyAI: a dashboard opt-out, paid plans only, and “If you are opted out of model training, we offer zero data retention of audio and transcripts for our Streaming product.” Check each vendor’s DPA, its subprocessor list and whether a BAA is needed. Deepgram’s pricing page signs BAAs for Enterprise customers; AssemblyAI’s Data Controls page offers one self-serve on paid plans. - Running the model yourself. Deepgram self-hosting is an Enterprise option, “in your own VPC or on-prem hardware” with NVIDIA GPUs, and Flux is available self-hosted. AssemblyAI’s self-hosted streaming “requires an upfront commercial commitment of $50,000”.
Which one should you pick?
The table puts the platform answer first, then the two APIs. It is a fit guide, not a scored ranking: the Deepgram and AssemblyAI rows come from their own pages, and the Zian and LeadsNow.ai rows from what those companies publish. zian.ai publishes this page and appears in it.
| If you are… | Pick | Why | Wrong for you if… |
|---|---|---|---|
| A business that wants calls answered and appointments booked, not an STT choice | Zian AI (platform); LeadsNow.ai for done-for-you | AI receptionist and outbound agents in 30+ languages with SMS, email, WhatsApp, Instagram DM and Messenger follow-up; LeadsNow.ai runs it for you, pay-per-result | You need to write your own agent code, or need a published price today |
| An English-only agent with no keyterm list | Deepgram Flux English | $0.0065 a minute while the promotion runs, model-integrated turn detection, Australian endpoint | You need keyterms (then AssemblyAI is cheaper) or the promotion has ended |
| An agent that needs keyterms, or callers in languages outside Flux’s 10 | AssemblyAI STT-3.7-Realtime-Agent | $0.45 an hour with keyterms included, turn detection across 32 languages, entity holds for numbers and addresses | Your audio must be processed in Australia or India without a $50,000 self-hosted commitment, or your code may leave sessions open |
| A team with its own turn detector and a long-tail language | Deepgram Nova-3 streaming | 99 languages, the lowest rate here at $0.0048 promotional, Configure for mid-stream keyterms | You want the engine to own the turn |
| A regulated team that must run speech in its own environment | Deepgram self-hosted (Enterprise) or AssemblyAI self-hosted ($50,000 commitment) | Audio stays on your infrastructure | You have no Enterprise budget or GPU operations team |
If you want the STT bundled with a model and voice behind one connection, our Vapi vs Deepgram Voice Agent API cost breakdown prices that route.
The sentence on each vendor’s pages that cuts against this page
On Deepgram’s side it is the badge on its own price table: “Promo pricing”. The Flux and Nova-3 rates that win the cheaper columns above are promotional, the page data ends the promotion on 13 February 2027, and the pricing page does not print what replaces them (we found no speech-to-text promotion terms page in deepgram.com’s sitemap as at 22 November 2026); at the “was” prices, Flux English costs more than AssemblyAI’s Agent model with or without keyterms. Flux also drops features Nova-3 has: Deepgram’s comparison marks Smart Formatting, Speaker Diarization and pre-recorded audio unsupported on Flux.
On AssemblyAI’s side it is the billing warning: “Streaming is billed per session, not per second of audio”, and “If a session is not closed, it will automatically close after 3 hours — and you’ll be billed for the full 3 hours of session time”. AssemblyAI says improperly closed sessions “are the most common cause of unexpected charges that lead to negative account balances”. Its flagship agent model is also four days old as at 22 November 2026, and the changelog entry that launched it says its notetaker sibling, AssemblyAI-STT-3.7-Realtime, replaces Universal-Streaming and Universal-Streaming Multilingual. Neither vendor’s accuracy chart is reproduced here: AssemblyAI’s home page shows itself first on a voice-agent entity-error chart it labels “Ranked first by independent benchmarks”, and Deepgram’s models page claims a 54.2% reduction in streaming word error rate “compared to competitors”. Test both on your own recorded calls.
Frequently asked questions
Is Deepgram cheaper than AssemblyAI for streaming?
For a turn-taking agent, only on English calls without keyterms while Deepgram’s promotion runs (its pricing page data ends it on 13 February 2027). Flux English is $0.0065 a minute against AssemblyAI-STT-3.7-Realtime-Agent at $0.0075, so 10,000 minutes cost $65.00 against $75.00. Deepgram’s $0.0013 keyterm add-on makes Flux $78.00, and AssemblyAI includes keyterms. Nova-3 Monolingual is cheaper either way at its promotional $0.0048 ($48.00, or $61.00 with keyterms), but it has no turn detection.
Which AssemblyAI model should a voice agent use?
AssemblyAI recommends AssemblyAI-STT-3.7-Realtime-Agent for voice agents, at $0.45 an hour. It is the only AssemblyAI model with a manual turn-detection mode. AssemblyAI-STT-3.7-Realtime costs $0.30 an hour but segments transcripts for notetakers rather than driving turn-taking.
How many languages does each support in streaming?
AssemblyAI’s streaming models cover 32 languages, all with turn detection and code-switching. Deepgram Flux takes turns in 10 languages, English included, while Nova-3 streams 99 languages without turn detection, 55 of them in one multilingual stream. Each count comes from the vendor’s own language page.
Do Deepgram and AssemblyAI accept 8 kHz phone audio?
Yes. Both accept raw 8 kHz mu-law, the format Twilio Media Streams send, with no resampling. On Deepgram set encoding to mulaw and sample_rate to 8000; on AssemblyAI set encoding to pcm_mulaw and sample_rate to 8000, which its WebSocket reference allows.
How does AssemblyAI bill streaming?
By session time, not audio. AssemblyAI’s billing page says streaming is billed on how long the WebSocket stays open, idle time included, and that an unterminated session closes after 3 hours and bills all 3. Deepgram bills per second of processed audio.
Can I process the audio in Australia?
On Deepgram, yes: its Australian endpoint keeps requests in Australia and fails rather than falling back to another region. AssemblyAI says its cloud streaming audio is processed in the United States and the European Union only; its self-hosted streaming, which requires a $50,000 commitment, keeps audio on your own infrastructure.
What is the Keyterm Swing Rule?
On English calls, Deepgram Flux at $0.0065 a minute is cheaper than AssemblyAI at $0.0075 only without keyterms. Deepgram’s $0.0013 keyterm add-on lifts Flux to $0.0078, above AssemblyAI, which includes keyterms. At Deepgram’s $0.0077 list rate, Flux costs more either way.
Where every figure on this page comes from
| Figure | Who published it | Link | Date read |
|---|---|---|---|
| Flux English $0.0065 (was $0.0077), Flux Multilingual $0.0078, Nova-3 Monolingual $0.0048 (was $0.0077), Nova-3 Multilingual $0.0058 (was $0.0092) a minute; Growth $0.0057; “Promo pricing” and promotion end 13 February 2027 (page data); per-second billing; add-ons (Keyterm $0.0013, Redaction $0.0020, Entity Detection $0.0017, Diarization $0.0020); 150 / 225 concurrent streams; Growth $4,000 a year; self-hosted on Enterprise; BAA for Enterprise; “60+ languages” | Deepgram | deepgram.com/pricing | 22 November 2026 |
| Flux eot_threshold 0.5 to 1.0 (0.7), eager_eot_threshold 0.3 to 0.9, eot_timeout_ms 500 to 60,000 (5,000) | Deepgram | developers.deepgram.com/docs/flux/configuration | 22 November 2026 |
| Configure message: thresholds and keyterms mid-stream without reconnecting; raise eot_timeout_ms and eot_threshold for “a password, OTP, or account number” | Deepgram | developers.deepgram.com/docs/flux/configure | 22 November 2026 |
| Eager mode “can increase LLM API calls by 50-70%” | Deepgram | developers.deepgram.com/docs/flux/quickstart | 22 November 2026 |
| Flux unsupported: Smart Formatting, Diarization, pre-recorded audio; 10-language flux-general-multi; Nova-3 multi 55 | Deepgram | developers.deepgram.com/docs/flux/flux-nova-3-comparison | 22 November 2026 |
| Nova-3 “no turn detection”; 99 single languages (100 rows); 55 multilingual; Flux language lists; 54.2% WER claim | Deepgram | developers.deepgram.com/docs/models-languages-overview | 22 November 2026 |
| 36 Nova-3 languages for batch and streaming (16 November 2026); Nova-3 multi 55, streaming and batch (23 October 2026) | Deepgram | developers.deepgram.com/changelog | 22 November 2026 |
| Keyterms up to 100, 500-token limit, Nova-3 and Flux | Deepgram | developers.deepgram.com/docs/keyterm | 22 November 2026 |
| Nova-3 endpointing default 10 ms | Deepgram | developers.deepgram.com/docs/endpointing | 22 November 2026 |
| mulaw at 8 kHz “with no resampling” | Deepgram | developers.deepgram.com/docs/twilio-and-deepgram-stt | 22 November 2026 |
| Flux encodings incl. mulaw and alaw | Deepgram | developers.deepgram.com/docs/encoding | 22 November 2026 |
| 150 / 225 streams per project; secondary projects single stream; streaming diarization 50 (North America) and 25 | Deepgram | developers.deepgram.com/reference/api-rate-limits | 22 November 2026 |
| EU, Australia, India endpoints; no fallback outside Australia | Deepgram | developers.deepgram.com/reference/regional-endpoints | 22 November 2026 |
| mip_opt_out; retained only for the duration of the request | Deepgram | developers.deepgram.com/docs/the-deepgram-model-improvement-partnership-program | 22 November 2026 |
| STT-3.7-Realtime-Agent $0.45, STT-3.7-Realtime $0.30 an hour; keyterms included (100, maximum 6 words); prompting $0.05, Voice Focus $0.10, diarization $0.12, PII redaction $0.12, Medical Mode $0.15; 100 new sessions a minute (FAQ); home-page benchmark | AssemblyAI | assemblyai.com/pricing | 22 November 2026 |
| Universal-3.6 Pro Realtime $0.45 an hour; streaming billed on session duration | AssemblyAI | assemblyai.com/docs/getting-started/models | 22 November 2026 |
| Session-time billing; 3-hour auto-close billed in full | AssemblyAI | assemblyai.com/docs/billing-and-pricing | 22 November 2026 |
| min/max turn silence by mode; entity holds 7.0% to 2.2%, p90 0.83 to 1.64 s; manual mode Agent only; STT-3.7-Realtime segments transcripts | AssemblyAI | assemblyai.com/docs/streaming/turn-detection | 22 November 2026 |
| Balanced is the default mode | AssemblyAI | assemblyai.com/docs/streaming/getting-started/optimizing-accuracy-and-latency | 22 November 2026 |
| 100 keyterms, 50 characters each | AssemblyAI | assemblyai.com/docs/streaming/prompting-and-keyterms | 22 November 2026 |
| 32 streaming languages, code-switching by default | AssemblyAI | assemblyai.com/docs/streaming/multilingual-transcription | 22 November 2026 |
| pcm_mulaw; sample_rate 8000 to 96000 | AssemblyAI | assemblyai.com/docs/streaming/api-spec/streaming-websocket | 22 November 2026 |
| [unclear] tagging tip for 8 kHz mulaw | AssemblyAI | assemblyai.com/docs/voice-agents/best-practices | 22 November 2026 |
| No cap on open sessions; 200 new sessions a minute; 10% scaling at 80% | AssemblyAI | assemblyai.com/docs/streaming/rate-limits | 22 November 2026 |
| US and EU processing only | AssemblyAI | assemblyai.com/docs/streaming/endpoints-and-data-zones | 22 November 2026 |
| Zero streaming retention when opted out | AssemblyAI | assemblyai.com/docs/data-retention-and-model-training | 22 November 2026 |
| Opt-out and BAA self-serve on paid plans | AssemblyAI | assemblyai.com/docs/data-controls | 22 November 2026 |
| $50,000 self-hosted commitment | AssemblyAI | assemblyai.com/docs/streaming/self-hosted-streaming | 22 November 2026 |
| STT-3.7 launch 18 November 2026; replaces Universal-Streaming and Universal-Streaming Multilingual | AssemblyAI | assemblyai.com/changelog | 22 November 2026 |
| Australian English among pre-recorded dialects | AssemblyAI | assemblyai.com/docs/pre-recorded-audio/supported-languages | 22 November 2026 |
| Worked bill, Keyterm Swing Rule, Growth break-even (about 58,480 minutes a month) | zian.ai calculation from the rates above | This page | 22 November 2026 |