What changed
On 2 November 2026, OpenAI's API changelog opened outbound SIP calling through the Live API to every organisation. SIP (Session Initiation Protocol) is the signalling standard phone carriers and VoIP systems use to set up calls. Until now, GPT-Live phone work was mostly inbound. With this release a voice agent built on gpt-live-1 can ring an ordinary phone number. Trade coverage dated the general availability announcement 4 November (TokenPost, Crypto Briefing). The changelog entry carries the earlier date.
OpenAI doesn't act as a carrier here. You bring a SIP trunk from a provider such as Twilio or Telnyx, and that provider carries the call onto the public phone network. GPT-Live handles the conversation. Crypto Briefing also lists Ozonetel as a supported provider.
How a GPT-Live call is put together
GPT-Live 1 is full duplex, meaning the model can listen and speak at once. The model handles speech and passes reasoning and tool use to a backend through a mechanism OpenAI calls delegation. The delegation guide describes two modes. In Responses delegation, GPT-Live calls a Responses API model you configure, and OpenAI suggests gpt-6-luna as a starting point. In client delegation, your own application receives each request and runs whatever model, agent or service it likes before sending results back. You choose the mode when you create the session and can't change it later.
For phone calls, your backend talks to the session over a "sideband" WebSocket. That's a separate control connection running alongside the audio, which your provider carries directly. According to the telephony guide, the sideband handles session events, call-progress events and call controls.
Building an outbound call, step by step
The steps below follow the OpenAI telephony and SIP guide.
- Prepare a secure trunk. GPT-Live requires TLS for signalling and SDES-SRTP for encrypted audio, and SRTP has to be enabled with your provider before you place a call. On Twilio, the setting is a switch on the Elastic SIP trunk in the Console called Secure Trunking, and Twilio's post says the feature costs nothing extra.
- Open the network paths. Allow outbound TCP/TLS on port 5061 to
sip.api.openai.com, or tosip-eu.api.openai.comif you need European data residency. Allow two-way UDP SRTP to the four/28ranges listed in the guide. - Create the session. Send a
POSTtohttps://api.openai.com/v1/live/sessionswith asessionobject and atransportobject of typesip. The guide builds the body withjqand sends it withcurl. The body follows this shape:
{
"session": {
"model": "gpt-live-1",
"instructions": "<your instructions>",
"delegation": { "type": "client" }
},
"transport": {
"type": "sip",
"destination": "<E.164 number>",
"trunk": {
"provider_url": "sips:<provider-host>:5061",
"auth": { "type": "digest", "username": "<user>", "password": "<pass>" },
"caller_number": "<E.164 caller number>"
}
}
}- Attach the sideband straight away. A
200 OKresponse means the session has initialised. The phone hasn't necessarily been answered. Take the session ID from the response and connect towss://api.openai.com/v1/live/sessions/{session_id}/attach. The sideband replays only the previous 3 seconds of events, so any delay means lost events. - Track progress. Watch for
transport.ringing,transport.answeredandtransport.failed, and deduplicate them byevent_id. - Control the call. Send keypad tones with a
transport.dtmf.sendevent, transfer withPOST /v1/live/sessions/{session_id}/referand atel:orsip:target, and end the call withPOST /v1/live/sessions/{session_id}/hangup. Keep the sideband open untilsession.closedarrives, because the final usage figures come in that event.
Leave audio.format out of the session, since SIP negotiates the audio format itself. The provider URL has to use sips:, and credentials belong in the auth block. Putting credentials in the URL isn't allowed, and the API rejects local or private hosts.
Feeding the agent results
When you use client delegation, the sideband sends a session.delegation.created event with a delegation ID and no task text. Your backend works out the task from the transcript deltas it has collected. The delegation guide offers three ways to return results. session.commentary.append sends content for GPT-Live to speak, session.thinking.append supplies background context for later replies, and session.instructions.append adds session-wide guidance. Each append accepts up to 500 tokens of plain text. The guide warns that transcript fragments can be incomplete or wrong. Treat short spoken replies such as "yes" or a read-out reference number as unconfirmed until your backend has checked them against task state.
Limits
The telephony guide caps ringing at 3 minutes and connected calls at 2 hours, and a request can't override either cap. Destinations have to be E.164 numbers, the international format that starts with a plus sign and country code (+64 for New Zealand). SIP URI destinations aren't supported. A failure while setting up the transport can return 502, and an initialisation timeout can return 504. Some third-party write-ups mention a 403 outbound_sip_not_enabled error. OpenAI's changelog says the feature is open to all organisations, so if you hit that error, raise it with OpenAI support.
The model page measures rate limits in concurrent sessions: 50 at the Build tier, 300 at Launch and 500 at Grow. The free tier isn't supported. GPT-Live handles audio and text only, and it runs solely on the v1/live/sessions endpoint.
Costs
Voice sessions cost US$0.05 a minute, billed per second, according to the model page. Backend model and tool usage is billed separately. Your SIP provider charges its own carrier fees on top. As Crypto Briefing puts it, the true cost of a call is "the sum of both bills". A ten-minute call therefore costs US$0.50 for the voice layer alone, plus backend tokens and whatever your carrier charges to terminate a call on a New Zealand mobile or landline. OpenAI's pages don't publish carrier rates, so get local termination pricing from your trunk provider before you size a campaign.
What it means for builders in New Zealand
The guide lists only two SIP hosts, a global host and an EU host. It offers no Australian or New Zealand residency option for signalling. If your project handles health, financial or other sensitive conversations, check where audio and transcripts are processed before you commit. OpenAI's documentation doesn't state what latency to expect from Auckland, so measure time to first audio and interruption handling on real local calls.
Sensible first builds include appointment reminders that use DTMF for confirmation, agents that work through supplier phone menus, and callback flows that transfer to a human through refer when the backend flags a problem. Client delegation suits teams that want to validate every result before the voice agent says it aloud. If you build something along these lines, we'd like to see it on our projects page.
Sources
- OpenAI API changelog
- OpenAI: Telephony and SIP guide (Live API)
- OpenAI: Delegation and tools in GPT-Live
- OpenAI: GPT-Live 1 model page
- Twilio: How to Secure Your SIP Trunks with Twilio
- Crypto Briefing: OpenAI launches GPT-Live outbound phone calls API at $0.05 per minute
- TokenPost: OpenAI Opens GPT-Live Calling API With $0.05 Voice-Layer Rate
Build with people working on the same things
Join the Auckland Society for AI: NZ$599 a year with a ForgeBox and the members' projects area. Our Auckland meetups are open to everyone.