AI Voice· Verdieping
GPT Live: from turn-taking to natural AI conversation
OpenAI brings full-duplex audio to the API. What does that mean for business voice agents, and what must you arrange before you start?
A voice agent that first waits until you have finished speaking, then thinks and afterwards speaks, feels different from a real conversation. With GPT‑Live‑1, OpenAI wants to break that sequence: the model can listen and speak simultaneously. OpenAI announced the API version on 10 September 2026. This article explains the confirmed capabilities and distinguishes them from our practical recommendations. We have not tested the model ourselves. [1]
What is GPT Live? GPT Live is OpenAI's system for continuous speech interaction. GPT‑Live‑1 is the voice layer that, according to OpenAI, is available via the API for developers. Full-duplex here means that incoming and outgoing audio are processed together: listening does not stop as soon as the agent begins to speak. The system can therefore respond to interruptions and brief confirmations whilst the conversation continues. [1][2]
In a classic chain, speech is first converted to text, processed by a language model and then spoken again. Each transfer can cause delay and loss of conversational signals. GPT Live processes listening and speaking in a single voice model. More complex reasoning and actions can be delegated to an underlying model or external tools. That does not automatically make the voice layer complete customer service: your own systems must still supply data and perform actions. [1]
More natural real-time audio: what does OpenAI confirm? OpenAI mentions interruption handling, adjustment of tone, tempo and conversational style via the system prompt, dealing with silence and background noise, and context retention in longer sessions. Telephony support, transcriptions and text of responses are also mentioned in the API announcement. These are product claims from OpenAI, not results from our own tests. [1]
The distinction is important: a pleasant-sounding voice does not prove that a business action has been carried out correctly. An agent can respond politely and still book the wrong appointment when your integration does not check sufficiently. Our recommendation is therefore to assess conversation quality and correct execution separately.
Which business applications make sense? OpenAI mentions, among others, restaurant reservations and customer service as telephony applications. The announcement contains examples from language teaching and customer support. These examples show the intended direction, but are no guarantee of the same results in your organisation. [1]
As editorial applications, we are thinking of appointment scheduling, order status queries and an initial inventory of a service request. Begin with tasks with clear boundaries: which information may be used, which action may the agent perform and when does an employee take over? Do not allow sensitive decisions, exceptions and unclear requests to be handled unnoticed by a voice agent. These are implementation recommendations, not additional capabilities that we attribute to the model.
Architecture and latency: not only the model counts The technical publication of 3 August describes a separate fast audio stream, with asynchronous handling of delegation and tool use. OpenAI uses WebRTC for media transport and describes how session context and model transitions are processed outside the critical audio path. [2]
According to OpenAI, that reduces blockages in the conversation. But a slow underlying system remains a problem: the technical source explicitly states that the voice layer cannot hide an arbitrarily slow response. A conversation can therefore continue whilst a tool is working, without the requested answer or action being ready yet. [2]
We give no millisecond promise for your application. Measure yourself the time to the first audible response and the time to a confirmed action, under realistic network conditions, interruptions and background noise. Also test what happens when a connection fails. We have not carried out these measurements.
How to prepare an initial pilot - Choose one limited task and establish when the agent must transfer. - Determine which tools may read information and which may execute changes. Request confirmation before an important action becomes final. - Keep API keys on the server and give tools only the necessary permissions. - Check the current API documentation for session configuration, supported connections and access conditions. The announcement is not a complete installation guide. - Test pauses, corrections, background noise and slowly responding systems; check the transcription and the actually executed action. - Set a cost limit and make transfer to a human simple.
This is our checklist for a responsible pilot, not a claim that GPT‑Live‑1 establishes these safeguards for you.
Privacy: a separate design decision The two consulted articles offer no complete description of retention periods, regional data processing or the privacy terms for your business deployment. We therefore do not conclude that conversations automatically remain in Europe, are not stored or comply with your GDPR obligations.
Determine in advance which personal data your voice agent needs, whether you want to retain audio and transcriptions, who gets access and when data are deleted. Check the current contractual terms and data processing with the supplier. Inform callers that they are speaking with AI and design an alternative for sensitive questions. These are editorial recommendations; have the concrete legal assessment for your organisation verified.
Availability and costs According to the announcement, GPT‑Live‑1 was available on 10 September 2026 in the API for $0.05 per minute for the front-end voice layer. That is not automatically the total conversation price: the underlying model, tools, telephony and your own infrastructure can cause separate costs. Check current rates and access before you budget a pilot. [1]
OpenAI describes expansion of voice options for accents, dialects and languages, but these two sources give no exhaustive list or specific guarantee for Dutch conversation quality. Therefore have the intended language and terminology assessed in your own pilot. [1]
Our conclusion GPT Live shifts the technical starting point: not first awaiting a turn to speak, but a continuous audio stream in which listening, speaking and outsourced reasoning can take place alongside each other. That is relevant for business voice agents. The value, however, only emerges when you couple that voice layer to reliable business processes, clear permissions, privacy choices and a measurable pilot. [1][2]
Primary sources - 1 — OpenAI: GPT‑Live‑1 in the API, 10 September 2026 - 2 — OpenAI: architecture for continuous voice interaction, 3 August 2026
AI-assisted draft by AI Up to Date Editorial based on the above primary sources. No own product test carried out.
Sources
- [1]OpenAI: Build more natural voice experiences with GPT‑Live‑1 in the API (10 September 2026)
- [2]OpenAI: How we built a realtime system for responsive voice AI in six months (3 August 2026)
This article was produced with AI assistance from the original sources cited and checked against our editorial policy. Editorial policy