Cover art for Speech-to-Speech Voice Agent Pricing and Architecture
Episode 6739 manual-script

Speech-to-Speech Voice Agent Pricing and Architecture

Gemini Live charges about half a cent per minute for audio input, while OpenAI bills $32 per million input tokens.

24:37 Transcript

0:00
0:00
Episode Details
Episode ID
MWP-6943
Published
Duration
24:37
Audio
Direct link
Pipeline
V5.36
TTS Engine
omnivoice
Script Writing Agent
GLM-5.3

AI-Generated Content: This podcast is created using AI personas. Please verify any important information independently.

The prompt

The prompt that started this episode, as sent in (production notes for the hosts removed).

Building on our earlier discussion about how speech-to-speech models are actually built: I want to size up the market for someone about to build a voice agent. I've built a few on Vapi with the cascade approach, and I'd always assumed the cascade was the thrifty option, so I was surprised when it came up on the show that speech-to-speech models can end up cheaper. Who are the ready-to-go speech-to-speech options right now? Deepgram is big in this space and Gemini has a voice agent SDK, but I mean the whole field. What is each one good at, and roughly what does it cost per minute of conversation?

Voice agent builders face a confusing market where "speech-to-speech" spans three distinct architectures under one label: native end-to-end audio models, managed orchestration layers, and pre-packaged transcription-to-speech cascades. While products like OpenAI's GPT Realtime family, Google's Gemini Live, and Amazon Nova 2 Sonic process raw audio directly, platforms like Deepgram Voice Agent API package traditional cascades under the same umbrella. Pricing models diverge just as sharply, splitting between flat connection fees and token-based billing that scales with call duration. Google challenges the market with aggressive token pricing at roughly half a cent per minute, whereas open-weight models like NVIDIA's VoiceChat showcase rapid gains in conversational naturalness but still lag behind proprietary leaders on complex reasoning benchmarks.

Mentions

  • Amazon Nova 2 Sonic Unified speech-to-speech model on Bedrock
  • Azure Voice Live Managed wrapper layer over selectable models
  • Deepgram ASR provider with text preprocessing guide
  • Gemini Live Google's multimodal speech-to-speech model
  • GPT Realtime OpenAI native audio model, token-priced
  • Grok Voice xAI flat-rate voice agent, OpenAI-Realtime-compatible socket
  • Moshi Kyutai open speech-to-speech model
  • NVIDIA VoiceChat Open-weight 11B full-duplex speech model
  • Ultravox Flat-rate speech model, MIT-licensed weights
  • Vapi Voice agent infrastructure platform

Downloads

Episode Audio

Download the full episode as an MP3 file

Download MP3
Transcript (TXT)

Plain text transcript file

Episode Book (PDF)

The episode's record — date, duration, models, sources — with the full transcript

#6739: Speech-to-Speech Voice Agent Pricing and Architecture

Corn
Last time we looked at how speech-to-speech models are actually built. This one is about what you can buy. Daniel has built voice agents on Vapi, the cascade way — transcribe, think, speak. And he'd always filed that under the thrifty option. Then we said on the show that native speech-to-speech can come out cheaper, and it bothered him enough to send this in.
Herman
Which is the correct response to being told something you've built several times is the expensive way of doing it.
Corn
So he wants the whole field. Not just Deepgram and the Gemini SDK he already knows about — everyone who's ready to go right now, what each one is actually good at, roughly what a minute costs, and the part he's really circling: why the cost assumption flips. All prices are as of November 2026.
Herman
And it's a better question than it looks, because before we name a single vendor, there's a trap to disarm. "Speech-to-speech" is not one product category. It's at least three wearing the same label, and the pricing models don't line up either.
Corn
Three architectures, at least three billing schemes. That's the mess we're walking into.
Herman
The first category is the real thing: native end-to-end audio models. One model, audio in, audio out, no transcription step in the middle. That's OpenAI's gpt-realtime family, Google's Gemini Live, Amazon Nova 2 Sonic, xAI's Grok Voice, NVIDIA's VoiceChat if you want open weights. Second category: managed layers. Microsoft's Azure Voice Live is the clearest case — it's not a model at all, it's a wrapper where you pick a model underneath and get telephony-grade extras around it. Third category, and this is the one that will matter to Daniel: managed cascades marketed as speech-to-speech. Products that are literally the speech-to-text, LLM, text-to-speech pipeline he already builds, wrapped up and sold under the speech-to-speech banner.
Corn
Same words on the tin, completely different machine inside.
Herman
And the pricing splits the same way — some are token-priced, which means your bill grows with the length of the call for reasons we'll get to, some are flat per-minute, some charge per minute of connection with your own components underneath, and one of the big ones we couldn't even get a dollar rate for.
Corn
The one where the price is a secret is the managed layer, I assume.
Herman
Azure Voice Live, yes. You can read the tier names all day, the dollars aren't on the page. Which is itself a data point.
Corn
It's a data point about their sales team. And the label survives because it describes the interface, not the internals: audio in, audio out looks the same from outside whether it's one model or a pipeline.
Herman
Let's start where the money is loudest. OpenAI, because they're running both plays at once, which almost nobody else is. GPT Realtime two point one is the native audio model, and it's priced per token — thirty-two dollars per million input audio tokens, sixty-four per million output, and cached input at forty cents.
Corn
Forty cents. From thirty-two.
Herman
That gap is doing an enormous amount of work, and we'll come back to it. The mini model is ten in, twenty out. Now the important bit for the arithmetic later: OpenAI counts user audio at one token per hundred milliseconds, which is six hundred tokens a minute, and assistant audio at one token per fifty milliseconds — twelve hundred a minute. The model talks twice as expensively as it listens. And every response resends the whole conversation.
Corn
The whole audio conversation. Every turn.
Herman
Remember that, because it's the hinge of the whole cost question later. Then in September, OpenAI launched GPT Live One, which is a different animal — a full-duplex voice layer at a flat five cents a minute, billed per second, with the reasoning model behind it billed separately. You can hang GPT six Astra off it, or a third-party model.
Corn
So they're selling the ears and the mouth, and renting you the brain by the hour.
Herman
That's a fair compression of it. They claim a thirty-point gain on Full Duplex Bench over GPT Realtime two point one — that's OpenAI's own claim, so treat it as marketing until someone independent runs it.
Corn
Google next, because this is the one that makes the cost conversation awkward for everyone.
Herman
Gemini Live. Two flavors on the API — the three point eight live model and the three point one flash live preview. Audio input at three dollars per million tokens, which works out to roughly half a cent a minute, and audio output at twelve per million, about one point eight cents. There's a free tier, which matters if you're prototyping. Feature-wise it's deep: native audio, voice activity detection built in, tool use, what they call affective dialog, proactive audio, twenty-four languages, thirty voices. And the SDK route is their Agent Development Kit streaming, which is the answer to the thing Daniel mentioned, the voice agent SDK.
Corn
Half a cent in, under two cents out. That's the number that makes the rest of the field uncomfortable.
Herman
It's the number that makes OpenAI's realtime pricing look like a different decade, yes. And proactive audio, where the model decides to speak up unprompted, is great for coaching or companion apps and a liability on a booking line. Amazon's Nova 2 Sonic sits in that same band — unified speech-to-speech on Bedrock, launched last December, a million-token context, twenty concurrent sessions per region by default. The speech tokens are reported at three dollars in, twelve out per million — that's secondary sources quoting the AWS table, I haven't seen it confirmed on AWS's own page, so hold it loosely. What it's good at is the AWS-shaped stuff: long context, and not having to leave Bedrock if you're already living there.
Corn
Which for a lot of enterprises is not a small thing. If your compliance posture already runs through AWS, "it's on Bedrock" beats "it's slightly better."
Herman
It beats a lot of things, in a lot of procurement meetings, yes.
Corn
Now the flat-rate one. This is where the pricing stops pretending to be about tokens.
Herman
Grok Voice. Five cents a minute at launch last December, and a third-party write-up citing xAI's October pricing page says eight cents now, no platform fee. I'd treat the eight as likely but not gospel. What I like about it structurally is the WebSocket — it's OpenAI-Realtime-compatible, so if you've built against OpenAI's realtime API, the port is nearly free. Server-side voice detection, tool use including web and X search, voice cloning from a clip up to two minutes. They claimed ninety-five percent on Big Bench Audio at launch — again, vendor number, their own benchmark show.
Corn
Cloning a voice from a two-minute clip is the kind of feature I'd use exactly once, catastrophically.
Herman
The OpenAI-compatible socket is the quietly big deal in that list. If a second vendor speaks the same protocol, you can A/B them on the same codebase, and price competition starts to actually function.
Corn
The open-weight side of native, briefly.
Herman
NVIDIA's VoiceChat, eleven billion parameters, full-duplex, tool calling, out since mid-year. On Big Bench Audio it scores about thirty-nine percent. Moshi, Kyutai's speech-to-speech model, scores four point three. The proprietary leaders are above ninety.
Corn
That's not a gap, that's a canyon.
Herman
It is. The interesting open models are catching up on voice naturalness much faster than on reasoning — VoiceChat is nine times Moshi's score and still less than half the proprietary floor. Naturalness you can demo in a thirty-second clip. Reasoning you find out about in week three of production.
Corn
One thing to keep straight if Daniel goes digging: Kyutai's Unmute, which gets mentioned in the same breath, is not a speech-to-speech model — it's a wrapper that puts open speech-to-text and text-to-speech around any text LLM. It's a cascade with good publicity.
Corn
Like the next one on the list, as it happens.
Herman
Which one's that?
Corn
Deepgram. He named them first, so let's give them the full reveal.
Herman
Deepgram Voice Agent API. And here's the reveal: it's a managed cascade. Flux or Nova-3 speech-to-text, then an LLM — Deepgram-managed or bring-your-own, and the bring-your-own list is OpenAI, Anthropic, Google, Groq, Bedrock, NVIDIA, plus their own think-fast models — then Aura-2 text-to-speech, or your own. What the API actually owns is end-of-thought detection, barge-in, function calling. The plumbing, not the speaking. They call it speech-to-speech in the audio-in-audio-out sense, and that's true of the interface, not the machinery.
Corn
So Daniel has been building a Deepgram-shaped thing by hand, and Deepgram is selling him the Deepgram-shaped thing pre-assembled, under the name of the thing it isn't.
Herman
That's the category confusion in one sentence. Their pricing is per minute of connection time: Standard at seven and a half cents pay-as-you-go, six point eight on the Growth plan; bring your own LLM and voice and it drops to five cents, four on Growth; Advanced at about sixteen cents. Note the unit, connection time — that's not the same as speech time, and we'll meet that distinction again.
Corn
It's per minute the phone line is open, not per minute anyone says anything.
Herman
And to be fair, the plumbing they own is the fiddly part of a hand-built cascade. The product is real; it's the label that misleads.
Corn
Then the managed layers. Azure Voice Live — you pick a model and a tier: Pro gets you GPT Realtime two point one and the GPT five family, Standard the minis, Lite has Microsoft's own Realtime Flash and Phi four multimodal in preview. Around it: noise suppression, echo cancellation, a hundred and forty speech-recognition locales, avatars. The wrinkle is that if you pick a non-realtime model, it runs through Azure's speech-to-text and text-to-speech — as a cascade. So even inside one product you can be on either architecture depending on the dropdown. And as established, no dollar rate on the page.
Herman
Which means the honest answer to "what does Azure Voice Live cost" is "I couldn't get a figure," and the plain answer to "is it native" is "check the dropdown." A product that can be either architecture depending on configuration is the taxonomy's reductio ad absurdum.
Corn
The one remaining tier is the small flat-rate specialists. Give me the quick pass.
Herman
Ultravox: five cents a minute, thirty free minutes, five concurrent calls on the free lane, a hundred a month for Pro. Audio goes straight into their LLM, no separate STT stage, MIT-licensed weights — but the older model cards say the output is text with a TTS step after, so it's only partly end-to-end. Hume, the emotion-aware one: third parties report four to seven cents a minute for EVI 3 depending on plan, about half that for the mini — Hume's own page didn't confirm, so those are hearsay numbers. And self-hosting open weights, VoiceChat or Moshi, needs a GPU, so your per-minute cost is entirely a question of utilization — no honest flat figure exists.
Corn
Say more about that last one, because "it depends on utilization" is the kind of answer that sounds like a shrug.
Herman
It's arithmetic, not a shrug. A GPU hour is a fixed cost whether it's handling one call or twenty. If you've got a card doing eight concurrent sessions around the clock, the per-minute cost is the card divided by the minutes; if it's handling one call for an hour a day, the same card costs you hundreds of times more per minute. Every self-hosting estimate you've ever seen that looked too good or too bad was just assuming a different fill rate.
Corn
That's the field. Now the part where his surprise makes sense: what a minute actually costs, and why the same question has two answers.
Herman
Let's do OpenAI's token math honestly, because this is where most people stop. Take a standard conversational minute: thirty seconds of user speech, twenty-five seconds of agent speech. User audio at six hundred tokens a minute, half a minute is three hundred tokens; assistant audio at twelve hundred a minute, twenty-five seconds is five hundred tokens. At thirty-two and sixty-four dollars per million: input about a cent, output about three point two cents.
Corn
So roughly four cents a minute, before anything interesting happens.
Herman
Before the interesting part, yes. Third-party planning ranges — this is Fora Soft's estimate, not OpenAI's — put GPT Realtime two point one at six to eleven cents a minute in practice, and long uncached calls much higher. The mini comes out near one point three cents before replay, with a realistic range of two to five. And Gemini Live, same speech mix, lands near one cent before text and context overhead.
Corn
And on the flat side: Live One at five cents plus the brain, Grok Voice five to eight, Ultravox five.
Herman
Now the other side of the ledger — what Daniel is actually paying. Vapi's own pricing page: five cents a minute hosting fee, Deepgram transcription at about one cent, an OpenAI LLM at somewhere between point eight and four and a half cents depending on the model, ElevenLabs voice at one and a half to two and a half cents.
Corn
Which totals out around eight to thirteen cents a minute. With the platform fee alone equal to Grok Voice's launch price.
Herman
That's the sentence I wanted to land. Vapi lists no speech-to-speech provider on that page — there's no native option sitting next to the cascade parts, so the comparison Daniel's been running in his head never even appears on the invoice. And the bare cascade, if you assemble it yourself with no platform — Deepgram Nova-3 at point seven seven cents a minute, a fast LLM, ElevenLabs' Flash voice — estimates put it at two to four cents. In that same cost model, text-to-speech is typically seventy percent or more of the component bill.
Corn
The mouth is the expensive part.
Herman
So here's where the surprise resolves, and I want to be careful with it, because the temptation is to declare a winner and it doesn't have one. "Speech-to-speech is cheaper" is true in some comparisons and false in others. Flagship token-priced native — GPT Realtime two point one — is two to four times a bare self-assembled cascade. Daniel's instinct is right there. But Gemini Live, the minis, and the flat-rate options at five to eight cents undercut the Vapi all-in bill of eight to thirteen cents, because that bill is carrying a five-cent platform fee before a single word is transcribed.
Corn
So he wasn't comparing architectures. He was comparing a bundled cascade to an unbundled native, or a bare cascade to a flagship native — and each pair flips the other way.
Herman
That's the rule. Not "cascade is cheap" and not "native is cheap" — which pair are you holding.
Corn
And the structural reason goes deeper than the fee. Tell me about the replay, because that's the mechanism I think most builders never see.
Herman
Every turn, the native model resends the whole conversation. Every response includes the audio history — your speech tokens, its speech tokens, all of it, again. A token-priced speech-to-speech call gets more expensive per minute the longer it runs, because minute ten is paying for minutes one through nine as baggage.
Corn
It's an actor who rereads the whole script every time he gets a line.
Herman
A text cascade resends history too — that's just how the context window works — but it's resending text, which is cheap, instead of audio tokens, which cost eight times the text input rate on OpenAI's sheet. Ten minutes of transcribed words is maybe two thousand tokens. The same ten minutes on GPT Realtime, at our thirty-and-twenty-five-second mix, is around eight thousand audio tokens, a quarter of a dollar per turn if nothing is cached.
Corn
So short calls favor token-priced native, long calls favor cascade, and the crossover is somewhere in between.
Herman
And the crossover moves depending on two things: how much history you're replaying, and whether caching is in play. That cached input price — forty cents per million against thirty-two dollars — is the lever. If your history is cached, replay mostly stops hurting; if it isn't, a long call on a flagship native model can run away on you.
Corn
That's the trap for someone building on Vapi, though. The five-cent platform fee dominates the cheap end of the cascade. The components are pennies — a cent for transcription, a couple for the voice — and then the platform is half the bill.
Herman
Which is exactly why a flat-rate native at five to eight cents can look cheaper than Daniel's Vapi bill even though the raw cascade underneath is nearly free by comparison. And it's why the two questions are entangled: build versus buy, and cascade versus native. You can't answer one without the other, because the platform fee only exists on one path and the replay penalty only exists on the other.
Larry
Friends, are you tired of your cascade dripping tokens while your wallet goes full duplex? Then you need the Sonic Cascade Whisper Bucket, the only speech-to-speech solution that is also a bucket. It wraps your native audio in a managed layer of genuine imitation leather.
Larry
Now here is the beautiful part. The price started at four hundred and ninety nine dollars, but while I have been talking it has dropped to two hundred and twelve, and now it is thirty one dollars and falling, and I am throwing in a per minute connection fee of absolutely nothing, a free voice activity detector, twenty concurrent sessions of my cousin Gary, and a limited edition token jar, but you have to act now because the deal just ended, and so has the episode, goodbye.
Herman
I own one of those, by the way. The bucket. It sits by the desk and the cats have claimed it, so it does nothing, but it does nothing very reliably.
Corn
It's the only product we've ever mentioned where the price falls fast enough to catch up with its own refund policy. Anyway — the fee structure. If Daniel wanted to act on any of this, what's the actual move?
Herman
Depends on his call lengths, and that's not a dodge. If his agents run short interactions — confirmations, lookups, quick triage — the flat-rate natives at five cents are simple and the replay penalty barely engages. If he's running long consultative calls, the token-priced natives are the ones to fear, and a lean cascade starts winning again the longer the call goes.
Corn
And there's a middle case, presumably — a call that starts short and gets long. A triage that turns into troubleshooting.
Herman
Which is the worst case for a token-priced native, because you're paying flagship replay rates on a call that outgrew its architecture mid-flight. If your call distribution has a fat tail, that tail is where the bill lives, and it's the part nobody models when they price off the average.
Corn
The numbers he'd want to watch aren't the sticker per-minute, either. It's the bill per minute on his actual call mix.
Herman
Right — his Vapi history is the dataset. He has real transcripts, real durations. He could take a month of calls, run the arithmetic both ways, and know his own crossover point exactly, instead of comparing list prices that measure different things. Token prices measure speech; Deepgram's is per minute of connection, so the gap opens up the moment the caller hesitates.
Hilbert
It's per minute of connection, and the two of you keep saying per minute like it's one number. Herman, you said it three times this episode and I let it go the first two.
Herman
Guilty. Corn was circling the same ditch, I think.
Corn
I was admiring it from the rim.
Hilbert
A quoted rate that turns out to be per connection-minute instead of per spoken-minute — I've had that argument with a supplier before. There was a client, a Mr. Feld at the garden centre on Rechovet, who was quoted a rate for an automated phone line. Four cents, he was told. His first bill was triple what he'd worked out on paper, and when he complained, it came out the four cents ran from connect to disconnect, and his customers spent most of the call deciding things. Silence, hold, a man reading his lottery numbers back to himself. All of it billed.
Herman
So the unit gap was the entire dispute.
Hilbert
The entire dispute. I keep a notebook of quoted rates against what actually showed up on invoices — it's mostly a few phone lines and some studio equipment, going back longer than it should — and Mr. Feld's was not the worst discrepancy in it. His was third.
Herman
Third out of how many, if the notebook's that specific?
Hilbert
Enough that I stopped counting them separately from the lines themselves. The thing your episode missed is that everything you're calling per-minute pricing bills two different clocks, and the clock nobody prices is the one where the caller is standing there thinking. Corn sits and thinks on air — you've heard him do it — imagine him on a voice line paying by the second.
Corn
I'd bankrupt a per-token model single-handedly.
Hilbert
You nearly bankrupted a per-connection one once, but that was a phone card and it's not on today's topic. Mr. Feld solved his problem the plain way, in the end — he put a forty-second cap on the call and told his customers to hang up and ring back if they needed longer. Two short calls cost him less than one long one, same amount of talking. He thought that was clever. I thought it was the only man in the argument who'd read his own bill, but I didn't tell him that.
Corn
So the workaround was making the customer do the session management.
Hilbert
The customer's cheaper than a developer. That's the part nobody puts in the pricing table.
Herman
The cap-and-callback trick is an answer to the replay problem too — it resets the history. A fresh call with a short context beats a long call dragging all its audio history along.
Hilbert
Mr. Feld didn't know that. He just knew what the envelope said. But you're right, and I'll grant it looked like wisdom by accident, which is how most of the wisdom I've seen arrives.
Herman
If the flat-rate natives keep multiplying — and Grok Voice's no-platform-fee model points that way — and the open-weight models close even half the reasoning gap, the cascade's cost advantage narrows down to the bare self-assembled case. Which is the one most builders on a platform never actually see. Daniel's five-cent hosting fee is the wall between him and the cheap version of his own architecture.
Corn
And the number he should watch on any of them is the bill per minute of his real calls, silence included.

This episode was generated with AI assistance. Hosts Herman and Corn are AI personalities. It may contain errors: disclaimer · report an error