GPT-Live-1 Splits Your Voice Agent In Two, And Your Paperwork Assumes One

OpenAI put GPT-Live-1 into the API on 10 September 2026, priced at five cents per minute, and the interesting part is not the number. GPT-Live-1 is a full duplex voice model, meaning it processes incoming audio while it is still producing outgoing audio, so it reacts to pauses, acknowledgements and interruptions continuously rather than waiting for a turn to end. That replaces the three stage pipeline almost every voice product has been built on since 2023, where speech to text feeds a language model which feeds text to speech, and where the seams between those three stages are exactly where the awkward silences live. The vendor reported numbers are a thirty percentage point improvement over GPT-Realtime-2.1 on Full Duplex Bench, and first place on Tau3 when paired with GPT-6 Astra at medium reasoning. Treat both as vendor evaluations until you have measured latency, recognition quality and tool reliability on your own traffic, because voice quality is unusually sensitive to accent, line quality and domain vocabulary, and none of those appear in a benchmark.

The architecture is the story, and it is easy to skim past because it reads like an implementation detail. GPT-Live-1 is a two layer design. The conversation layer, which is GPT-Live-1 itself, manages timing, barge in and the feel of the call. The reasoning and action layer is a separate backend model that you choose and configure, and it handles logic, tool calls and API work asynchronously while the voice layer keeps the caller company. That decoupling is genuinely clever engineering, and it is the reason the per minute rate can be low: the expensive thinking has been moved somewhere else. It also means that what your customer experiences as one agent is, inside your stack, two models from potentially two vendors, with two billing meters, two sets of logs and two failure modes. Every governance artefact you have, from the AI inventory to the data flow diagram to the incident runbook, almost certainly describes it as one thing.

Start with cost, because it is the trap that will catch teams fastest. Five cents a minute buys the front end voice session and nothing else. You then pay for the backend model tokens, and for whatever agent harness sits around it, and those are billed in a completely different unit. Elapsed time and tokens do not move together. A caller who thinks slowly, or who is put on hold while a tool call resolves, costs you minutes without generating tokens. A caller who asks one dense question costs you tokens without generating minutes. We have made the point before that vendors increasingly define their own unit of account, and this is the sharpest example yet, because the two units in a single conversation are not merely different prices, they are different physics. The routing decision that follows, cheap backend for routine calls and a stronger model for hard ones, is now simultaneously a quality control and a cost control, and there is no single dashboard anywhere that shows you both halves of the bill for one call.

The more useful document OpenAI shipped alongside the model is the list of things it says the application must provide: call routing, identity and authorisation controls, backend integrations, escalation to a human agent, audit logging, and explicit confirmation policies for consequential transactions. That is unusually candid, and it deserves to be read as a scope statement rather than a roadmap. Every item on that list is a control, and every one of them is yours. A platform that tells you plainly which controls it does not implement is doing you a favour, because the alternative is discovering the same list during an assessment. Print it, and put a named owner against each line before the first production call, because the gap between that list and what a working prototype actually has is where the next twelve months of voice agent incidents will come from.

Audit logging is the line on that list that looks solved and is not. GPT-Live-1 natively provides transcripts of incoming speech and the text of its own responses, which is real and useful and covers compliance, quality assurance and accessibility needs for the conversation itself. What it does not cover is what the agent did. The booking that was made, the refund that was issued, the record that was updated, the payment that was taken: all of that happened in the backend layer, on a separate clock, in a separate log, with no shared identifier joining it to the words that caused it unless you create one. This is the attribution gap we described when user interface operating agents started acting inside real systems, and the voice version is worse, because the evidence a regulator or a plaintiff will ask for is precisely the join. What was said, what was decided, and what was executed, in order, with timestamps that reconcile. Two well kept logs with no correlation identifier between them is not an audit trail. It is two thirds of an argument.

Disclosure gets harder rather than easier, which is a strange consequence of a better product. The transparency obligations in Article 50 of the EU AI Act became applicable on 2 August 2026, and they require that a person interacting with an AI system is told so. Full duplex means the caller can talk over the disclosure, and a model designed to yield gracefully to interruption will do exactly that. So the question becomes concrete and slightly absurd: if the caller interrupts at the third word, was the disclosure made? The disclosure now lives in the layer that is specifically engineered to be interruptible, which means it needs a retry rule, a log entry recording that it completed, and a decision about what happens on a call where it never did. The biometric question rides along with it, because the audio and the transcripts are voice data, and retention of that material is a decision somebody should make deliberately rather than inherit from a default. None of this is hard. It is just invisible until somebody asks, and by then the calls have already happened.

The framework mapping is mostly favourable, with one wrinkle worth flagging. Under ISO 42001 you have to decide whether this is one AI system with two components or two systems, and the honest answer is one system with two suppliers, which means the impact assessment covers the whole call while supplier management covers each layer separately. ISO 27001 supplier management now has two entries where it had one, and if the backend is a different provider from the voice layer, that is two subprocessor disclosures in your customer contracts, not one. SOC 2 will ask whether the human escalation path and the confirmation step for consequential transactions actually operated across a period, which is a question about logs rather than policy, and loops straight back to the correlation identifier. If the call ever touches card data then PCI DSS applies to the audio and the transcript as well as the payment call, and full duplex audio is a recording whether or not you think of it as one. Vanta, Drata, Secureframe, Sprinto, Thoropass and Hyperproof will all hold this register and schedule the reviews once somebody populates them, but none of them will notice on your behalf that one inventory row has quietly become two.

There is also a repricing effect that will move faster than any of the governance work, and it has an enumeration problem attached. A low per minute conversation layer with the reasoning unbundled puts direct pressure on the existing voice stack, so expect the ElevenLabs, Murf and Play.ht tier of the market to respond on price and on agent features within the quarter, and expect a wave of teams to rebuild working voice products on the new shape simply because the demo is better. Some of those teams will not be engineering teams. An operations manager can wire a booking line together in Bolt, v0 or Lovable, point it at a calendar and a payment provider, and have something that takes real calls from real customers by Friday. That is the population that will never appear in the inventory, never write the confirmation policy, and never log the disclosure, and it is the same unregistered agent population we keep returning to. The tooling has changed. The enumeration problem has not.

The work this week is small if you do it before you ship and awkward if you do it after. Add the backend model to your AI inventory as its own row, with its own owner and its own supplier record, rather than letting it hide inside the voice entry. Put one correlation identifier across both layers now, while it costs an afternoon, because retrofitting a join across two log stores after six months of calls is a data project. Measure real cost over a week of your own traffic, counting pauses, hold time and escalated calls, instead of trusting per minute arithmetic against an average call length nobody has verified. Write down where the disclosure is spoken, what the retry rule is when a caller interrupts it, and how you would prove it happened on a specific call from three weeks ago. Decide the retention period for audio and transcripts on purpose. And settle the confirmation policy for consequential transactions before a caller finds the gap for you, because the first person to discover that your agent will issue a refund without a second check will not be an auditor.