OpenAI's New Voice Model Will Not Promise to Say Your Disclosure Word for Word

September 14, 2026. If you sell or buy AI phone agents, the important sentence in OpenAI's new voice release is not the price. It is the one in the documentation that says the model cannot be relied on to say a line exactly as you wrote it. OpenAI made gpt-live-1 available in the API on 10 September, and its own docs state that asking the model to deliver specific wording "does not guarantee exact delivery" and that "if exact audio delivery is required, play a verified recording or rendered clip through your application and control GPT-Live output while it plays." Every US state that requires an AI caller to identify itself requires specific wording. So does every opt-out line you have ever been asked to script. That makes this a design constraint, not a footnote.

What OpenAI shipped on 10 September

The launch post is datelined September 10, 2026 and says GPT-Live 1 "is available in the API today at $0.05 per minute for the front-end voice layer." The short version of what is new:

1. A full duplex voice frontend. The model listens and speaks at the same time, and the conversation continues while a backend model or agent handles reasoning and tool calls.
2. Two delegation modes. Responses delegation, where GPT-Live calls an OpenAI model you choose, or client delegation, where your own application runs the workflow and sends results back.
3. A new endpoint. Sessions are created at v1/live/sessions. The model page marks v1/realtime as not supported for gpt-live-1, so this is a separate path, not a version bump.
4. Rate limits measured in concurrent sessions rather than tokens per minute.

On price, the model page is the rate card: "Voice sessions cost $0.05 per minute, billed per second. Backend model and tool usage is billed separately," and "session duration is not rounded up to the next whole minute." One detail worth knowing before you model a pilot: creating a WebRTC session "bills 15 seconds of voice duration while the session initializes," and that amount is credited back once the session runs, so a 90 second session is billed as 90 seconds and not 105.

The disclosure problem, in OpenAI's own words

A scripted disclosure is the most common compliance requirement in outbound and inbound calling, and it is the one thing a generative voice model is structurally bad at. OpenAI does not hide this. The guidance is that a prompt asking for particular wording does not guarantee the words come out that way, and that the remedy is to stop asking the model and play audio you control instead.

Read alongside the rest of the docs, that produces a concrete architecture. The disclosure is a file, not a prompt. Your application plays it, and it suppresses model output while the file plays. If you have been quoting AI calling work on the assumption that the script is the deliverable, the script is now two deliverables: the prompt, and the recorded lines that must never vary. Our note on TCPA compliant AI calling sets out which lines those usually are, because the wording is set by statute in several states and not by you.

Answering machine detection needs its own detector

The second constraint follows from the same architecture. GPT-Live "hears the incoming audio," but its voice frontend "delegates work instead of issuing ordinary structured function calls," and "a delegated backend does not automatically receive the waveform." In client mode the delegation event carries metadata and timing "without raw audio." So the component that decides whether a human or a voicemail greeting answered cannot be the model, unless you route audio to it deliberately.

OpenAI's migration guidance is to "explicitly route incoming audio to an audio-capable detector," and it offers, as "one application managed architecture to evaluate," running a separate Realtime session alongside GPT-Live for part of the call with a copy of the incoming audio sent to both. That is an option the vendor suggests you test, not a recommendation, and the docs tell you to choose when to stop the detector based on "its added cost." Two further cautions in the same guidance are worth pinning to the wall: recognising voicemail "does not establish that the greeting and beep have finished or that recording can begin," and there is no event marking the end of each spoken response, so playback has to be tracked in your client.

It answers calls. It does not place them.

The SIP documentation describes an inbound flow, states plainly that "this flow accepts inbound calls," and closes with the line that matters to anyone selling outbound: creating an outbound SIP call through the sessions endpoint "is not supported," and you should "use the relevant partner integration for provider owned outbound calling." Read that narrowly, because it is narrow. Origination stays where it already lives, with your telephony provider. OpenAI's partner documentation describes Twilio as connecting "incoming and outgoing phone calls to GPT-Live with Twilio Agent Connect" and Telnyx as a way to "build outbound calling experiences with GPT-Live and the Telnyx Voice API," and the SIP page documents a second, non partner path where your application bridges provider audio over a WebSocket. Outbound is available. It is simply not a property of this endpoint. One more gate: the docs tell you to "confirm that GPT-Live SIP support is enabled for your project," so SIP is switched on per project rather than by default.

Concurrency is the ceiling, not price

Rate limits are published as concurrent sessions: 25 at tier 1, 50 at tier 2, 200 at tier 3, 300 at tier 4 and 500 at tier 5, with the free tier listed as unsupported. For anyone sizing a call centre this is the number that binds. Twenty five simultaneous calls is a small queue, and usage tier is a function of billing history, so capacity is something you grow into rather than buy on day one.

Three other published limits shape the build: instructions are capped at 16,384 tokens and cannot change after startup, startup history is capped at 128 messages and 8,192 tokens, and recordings expire after 30 days. The voice is also fixed once a session begins, and attempting to change it returns an immutable_field_update error. There are 12 newly listed voices with marin as the default, of which ten are English and two, Bossa and Tempo, are Brazilian Portuguese. Custom voices exist but are limited to eligible customers, capped at 20 per organisation, and documented as supporting English accents.

Two things not to conclude

First, there is no published session length limit. The docs acknowledge one exists, through a session closed reason of expired described as "the session reached its duration limit," but no value appears next to it, while the same pages publish precise numbers for instructions, history, context and retention. If you are quoting long calls, treat the cap as unknown and test for it rather than importing a figure from the Realtime API, which is a different product.

Second, GPT-Live does not retire the Realtime API. The deprecation in this area was the gpt-realtime model snapshot, notified on 20 July 2026 for removal on 20 January 2027, and its published replacement is gpt-realtime-2.1, another Realtime model. The Realtime API itself remains current and, as the answering machine guidance shows, OpenAI still expects some builds to run both.

What it means for operators

The practical read is that a voice agent is now assembled from four parts rather than one: a voice frontend you rent by the minute, a reasoning backend you pay for separately, a telephony provider that owns origination, and a set of recorded audio files that carry anything legally required. Quote all four or you will quote the first and absorb the rest.

If you are commissioning a build this quarter, four questions are worth putting in writing before anyone signs. Which lines in this call are required wording, and will they be played from a file rather than generated? Who places the outbound call, and on whose account? What is our concurrent session ceiling today, and what happens to call 26? And what detects voicemail, given that the model will not hand your backend the audio unless you make it. Those are the same questions we work through on AI automation builds, and they are cheaper to answer in a specification than in a post mortem.

One last item that belongs in a risk register rather than a sales deck: the session documentation notes that a moderation event can end a live session mid conversation, or cut off assistant audio without closing the session at all. A call that stops talking is a worse customer experience than a call that never connected, so decide now what your application does when the voice goes quiet.