Full-Duplex Speech-to-Speech AI Answers at $0.05 a Minute

Full-Duplex Speech-to-Speech AI Answers at $0.05 a Minute

OpenAI’s GPT-Live-1 charges $0.05 a minute for full-duplex speech-to-speech AI over a live voice session, billed by the second. And that number is why phone automation stopped being an enterprise-only line item. Full-duplex speech-to-speech AI for real-time phone calls is a single model that takes the caller’s audio in and sends audio back, with no transcript and no separate voice model in the middle, while continuing to listen as it speaks. A caller can cut in mid-sentence, say “wait, that’s the wrong address,” or drop an “mm-hm” while the model talks. And the model reacts instead of barreling ahead. Sessions run over WebRTC, WebSockets, telephony. And SIP, so the same system answers a browser demo and a real phone line. The architecture is genuinely new. The failure modes are not, and that distinction decides whether your build works.

No Transcript in the Middle Changes Everything

Older voice bots run a relay race: transcribe the caller, hand the text to a language model, hand the reply to a text-to-speech engine. Every handoff adds delay, and the transcript is a lossy copy of what the person actually said. Speech-to-speech cuts the relays. As Dograh’s writeup puts it, the model “takes the caller’s audio in and sends audio back, with no transcript and no separate voice model in the middle.”

The phone wiring is refreshingly boring.

Your carrier sends the call to your server, your server accepts it. And the AI provider handles the audio (Dograh). GPT-Live’s live sessions run over WebRTC, WebSockets, telephony. And SIP, and the model accepts audio and text in, and returns audio and text out (Retell). For a small operation, that means you point a phone number at your server and skip building a telephony stack yourself.

The design decision that matters most is the split. GPT-Live keeps the conversation in one model and “delegates reasoning and tool use to a backend agent,” per Dograh’s breakdown. Conversation handling and thinking are separate systems with separate costs, and you will feel that separation on the invoice.

Interruption Handling Beats Model IQ on the Phone

On a call, the failure that loses the customer is rarely a wrong answer. It’s the bot plowing through a correction because it finished “its turn.” Full-duplex systems are built to fix exactly that: the model keeps listening while it speaks. So a caller saying “mm-hm” or “wait” mid-sentence changes what happens next (Dograh). When both parties start talking at once, it “recovers when you both start at once,” as fullduplex.ai describes it.

My take: floor-taking discipline matters more than benchmark scores for phone work. GPT-Live handles interruptions itself rather than requiring you to bolt on an interruption-management layer, which is one less component you own and debug at 2 a.m. (Retell). Roark AI frames the constraint bluntly: “The only interface that reliably exists is audio in, audio out, over real time” (Roark). Everything else about your agent is negotiable. The audio stream is not.

For small businesses, the calls worth automating first are address corrections, reschedules, and order-status checks.

Those are precisely the calls where people interrupt, and precisely where turn-based bots fall apart.

What $0.05 a Minute Buys, and What It Doesn’t

Retell’s GPT-Live-1 writeup lists the voice session at $0.05 per minute, billed per second, with backend model and tool usage billed separately. Read that twice: two meters, not one. The voice minute is the floor. And every lookup or reasoning hop your backend agent makes rides on top of it.

Concurrency is tiered, and this is where a spike hurts you:

| Tier | Concurrent sessions |
|—|—|
| Build | 50 |
| Launch | 300 |
| Grow | 500 |

It is not available on the Free tier at all (Retell).

If Monday morning brings more than 50 simultaneous calls, you’re not on Build anymore.

On responsiveness, one comparison board reports Eleven Realtime v1 Max answering in 0.57 seconds and Grok Voice 3.0 in 0.81 seconds. That gap is the difference between a conversation and a long-distance call from 2003.

Two footnotes worth tracking.

Retell reports GPT-Live-1.1 rolled out to ChatGPT, Codex. And dots on October 15, 2026, announced as coming soon to the API. And not in the API as of November 5 (Retell). Build against 1 today, plan for 1.1. And if you want out of per-minute billing entirely, the open path exists: Kyutai’s Moshi, released in 2024, was the first openly released real-time full-duplex spoken LLM (Scaling Knowledge). And NVIDIA built PersonaPlex 7B on top of it, with a persona set from a voice sample and a text prompt (Dograh).

Whether self-hosting beats $0.05 a minute is your own ops math.

Your WebSocket Demo Is Lying to You

Here is the part that burns people.

Roark’s line deserves to be pinned above every voice project: “The model’s behavior on a WebSocket loopback is not its behavior on a PSTN call with jitter, packet loss, and codec compression” (Roark).

Real phone networks mangle audio in ways your laptop never will. A demo that feels flawless over WebSockets can turn into a stammering mess the moment it meets an actual carrier.

Test the way callers actually behave.

Roark’s simulation testing dials your agent over real phone calls, using personas that define the caller’s voice, language, accent, pace, emotional register, and background-noise environment (Roark).

Dograh’s pre-launch advice is simpler and free: “place calls where the caller corrects a detail or cuts in.

And read the traces” (Dograh).

The trap is that polite testing proves nothing.

Everyone tests with a cooperative caller on a quiet line.

Nobody who calls your business is that person.

What I’d Do Before Pointing a Number at It

Run the arithmetic on both meters first, voice minutes plus backend usage, so the first invoice doesn’t surprise you.

Then test on a real phone number, not a browser tab, with callers who interrupt, correct details. And talk over the model. Check your peak simultaneous calls against the concurrency tier you’re on before launch day. Watch for GPT-Live-1.1 hitting the API, since your pricing and behavior assumptions will shift the day it does.

Full-duplex speech-to-speech is the first voice tech that handles a phone call the way people actually hold one, at $0.05 a minute.

Pick the one call you’d automate first. And run it through the interruption gauntlet before you let it near a customer. If you’re weighing GPT-Live against Moshi for your own line, drop the use case in the comments and I’ll tell you where I think the money goes.

Leave a Reply

Your email address will not be published. Required fields are marked *