GPT-Live-1 voice agents, now on Open Voice Shield

Voice systems until now were turn-based. You talk, the system waits for silence, then it replies, and if you speak again too soon the two of you collide. GPT-Live-1 is full-duplex: it listens and speaks at the same time. On a phone call, that means it can:

• backchannel with a "mhmm" or a "yeah" while you are still talking;
• wait when you pause to think, instead of jumping in at the first gap;
• let you interrupt or change direction mid-sentence;
• keep talking while a backend model looks something up or calls a tool.

OpenAI and early users keep pointing at the same numbers: far fewer unwanted interruptions (the language-learning app Speak reported around 80% fewer), turn-taking around 0.8 seconds where the older Realtime model took about 1.4, and much better results on the conversational-dynamics benchmarks that measure exactly this. Those are their figures, not ours. On the calls we placed while building this, the agent answered a pause the way a person does.

The second change is architectural. GPT-Live-1 is the conversation layer, not the whole agent. It owns the live audio path. Deeper reasoning, search and tools are delegated to a backend model of your choice, and the conversation does not freeze while that work runs. That makes reservation agents, tutors and support lines practical on a real phone call, because the agent stays in the conversation while it works.

OpenAI exposes GPT-Live-1 over SIP, but not the way a carrier route works. A call to your project does not connect until a webhook of yours accepts it and tells OpenAI which model, voice and instructions to use. The agent only runs while a control connection is attached to the session, and the media has to be encrypted. Most switches cannot do any of that. The platform does all of it, per destination:

The webhook, hosted

Each OpenAI destination gets its own webhook URL. Register it in your OpenAI project, store the signing secret and a project API key on the destination, and the platform accepts every incoming call with the model, voice and instructions you set. Signatures are verified, redeliveries are handled once, and the destination shows the outcome of the last event.

The session, kept alive

After accepting, the platform holds OpenAI's control connection for the whole call. Without it the agent negotiates audio and then sits in silence.

Encryption where it is required, plain SIP where you are

The agent leg is TLS signalling and SRTP media, as OpenAI requires. Your switch keeps plain SIP and plain RTP and sees a normal call it can answer, acknowledge and hang up. We tested this against a carrier softswitch that ignores route sets entirely, and it works.

Recorded and analyzed like any call

Both legs are relayed through the platform's media relay and recorded in stereo. The call record carries the transcript, the verdict, the SIP ladder and per-leg quality, and the agent leg is metered per second.

The agent's settings live on the destination: the voice, the instructions, and the reasoning model GPT-Live-1 delegates to (the destination defaults to OpenAI's current recommendation). One lesson from testing: an agent told to "go along with the caller" waits for the caller to speak. If you want callers to hear it first, say so in the instructions, for example "Answer the call immediately by saying: Hello?".

1. Under Network, add a destination of kind Voice AI, provider OpenAI, with your project id as the Agent ID. Save it. The destination now shows its webhook URL with a copy button.
2. In your OpenAI project, under Settings → Webhooks, add that URL and subscribe it to the incoming-call event for GPT-Live. Copy the signing secret. Create a project API key in the same project.
3. Back on the destination, paste the signing secret and the API key, choose a voice and write the instructions. The "Needs webhook setup" flag clears.
4. Open the Web Dialer, pick the destination and call it. Or point a routing rule or a divert rule at it and let real traffic reach the agent.

The step-by-step version, with the API calls, is in the SIP integration guide. The Realtime family of models is supported on the same destination type. Subscribe the webhook to the Realtime incoming-call event instead and set a Realtime model on the destination.

• Voice AI destinations now support xAI and OpenAI, each with its own connection rules. The provider list is published by the API and in the portal.
• The agent's transfer, hangup-by-agent and tool workflows run on OpenAI's side. The platform relays the call and holds the session. Turning a transfer into a platform action is on the list.
• Beta. Exercised against test agents from a browser and from a carrier softswitch. The agent leg is metered per second at a rate you can see on the call record.

GPT-Live-1 is the first voice model you would put on a real phone line, because it stays in the conversation while it works. Open Voice Shield makes it a destination: one dialog, one webhook URL to register, and your existing switch or the web dialer does the rest, with every call recorded and analyzed.

Related: the voice agent gateway and the connect any SBC to a voice agent guide, xAI voice agents and G.722, divert known fraud callers to an AI voice agent, or start free with $5 credit.