GPT-Live-1 Makes Voice Cheap. Nobody Can Measure It Yet

GPT-Live-1 arrived in the OpenAI API on 10 September 2026 at $0.05 per minute, and it is the first voice model from OpenAI that listens and speaks at the same time. That capability is called full duplex, and it is the difference between a voice assistant that waits for you to finish and one that behaves like a person on a phone call.

The price is the headline for most developers. The interruption handling is the headline for anyone thinking about voice as a discovery channel.

What full duplex actually means

Older voice systems worked in turns. You spoke, it processed, it replied. If you interrupted, it either talked over you or lost the thread.

GPT-Live-1 listens while it is speaking. It can handle an interruption mid-sentence, register an acknowledgement like "mm-hm" without treating it as a new instruction, and sit through background noise or silence without narrating every step or filling the gap.

OpenAI first previewed the model in its GPT-Live announcement. It ships 12 voice options and supports custom voices, and the model is designed to work over telephony. The named use cases are reservations, order updates and customer service, which tells you where OpenAI expects the volume to be.

The architecture is worth understanding because it explains the price. GPT-Live-1 is the front-end voice layer. It holds the conversation, and it delegates the hard thinking to a separate backend model such as GPT-6 Astra or Codex, or to tools. So $0.05 per minute buys the voice, and you pay separately for whatever reasoning sits behind it.

Why the pricing structure is the real news

Splitting the voice layer from the reasoning layer is a design decision with a commercial consequence.

It means a company can run a cheap, fast, natural-sounding conversation on top of an expensive model that is only invoked when the caller actually needs something hard. Most of a support call is not hard. It is greeting, confirmation, clarification and closing. Paying frontier prices for that was the thing that kept voice agents uneconomic.

At five cents a minute for the conversational layer, a ten-minute call costs fifty cents to hold, plus whatever the reasoning costs. That is inside the range where replacing a queue becomes a finance decision rather than an innovation project.

This is the same pattern that has run through the whole model market: the floor collapses while the ceiling rises, and the collapse at the floor is what causes deployment. Cheap capability is what puts an assistant in places nobody would have paid to put one.

Voice as a discovery surface nobody is measuring

Here is the part that belongs on a marketing roadmap rather than an engineering one.

Every AI visibility conversation so far has assumed a screen. Citations are links. Links are things you click. The entire measurement apparatus, mine included, is built on the idea that being quoted eventually shows up as a visit.

A voice answer has no link. If a caller asks an assistant to book a table, compare two suppliers or check whether a shop has something in stock, the brand is either named in the spoken answer or it does not exist for that customer. There is no position two, and unlike a results page, there is no visible source list to audit afterwards.

That makes voice the least measurable surface in AI search, and it is arriving with a price tag designed for volume. I have written about why citation share is not traffic. Voice is the extreme version of that gap: influence with no artefact at all.

What to do about it

Four things, in rough order of urgency.

Make your core facts speakable. Opening hours, availability, price, location, returns policy, lead time. Short, complete, unambiguous sentences that a model can read aloud without hedging. A paragraph that needs a table to make sense cannot be spoken.

Get your structured data right, because it is what voice reaches for. Business hours, location, product availability and pricing in machine-readable form are no longer a rich-result nicety. They are the raw material of a spoken answer.

Test your own brand by voice. Ask an assistant the ten questions your customers ring you about. Listen to what it says about you. Most teams have never done this once, and it takes an afternoon.

Decide what a voice interaction is worth before you can measure it. You will not get clean attribution here for years. Agree internally what being the named answer is worth, so the absence of a click does not become an argument for defunding the work.

The caveat on the numbers

One honest note. The $0.05 per minute figure and the 12 voices come from OpenAI's launch materials and the reporting around them, not from a bill I have paid. The effective cost of a real deployment depends entirely on the backend model you pair with it and how often it gets invoked, and OpenAI has published no benchmark for how frequently the front-end layer can answer alone.

Anyone modelling this should build the estimate on their own call mix rather than the headline rate. A support line where most calls need a lookup will not cost five cents a minute.

What GPT-Live-1 changes for search and discovery teams

The short version: a new discovery surface just became cheap, and it is one where your visibility cannot be tracked by any tool you currently own.

That is not a reason to panic, and it is certainly not a reason to buy a voice strategy from anyone selling one this month. It is a reason to make the facts about your business short, current, structured and quotable, because that is the only preparation that pays off across every surface at once, spoken or otherwise. GPT-Live-1 makes voice economic. What it says about you is still decided by what you have published.