OpenAI Adds SIP Calling to GPT-Live: Voice AI at 5 Cents a Minute

Until now, for an AI agent to call a customer or take a call, you needed a separate layer between the phone network and the model: a telephony service, a voice orchestrator, and sometimes your own media server. OpenAI is taking a step toward making that layer thinner: the company has published a guide to connecting GPT-Live via SIP, the standard IP telephony protocol.

How it connects

The setup is fairly straightforward. For incoming calls, you get a number from any SIP provider. The author of the review mentions Zadarma, but any other will do. Then you configure a webhook for the live.transport.incoming event and route traffic to sip.api.openai.com. For European customers there is a separate sip-eu address, which matters if you care about latency and about where data is processed.

Outgoing calls are started with a request to /v1/live/sessions, which specifies where to call. From there, the model conducts the conversation.

The platform includes basic telephone functions: transferring a call to another person, sending digits (the familiar "press 1" in a menu), and ending the call. There are also ready-made integrations with LiveKit, Twilio, Telnyx, Bandwidth, and Daily/Pipecat. In other words, if you already work with one of these services, you don't have to switch to the new setup.

Price: 5 cents per minute, billed per second. But that covers only the telephony and voice layer; the model and tools are paid for separately. You should calculate the cost of a conversation as the sum of all components, not from this figure alone.

What remains on your side

The author's main conclusion: out of the box, this is still a raw construction kit, not a finished product for outbound calling. A lot you will have to implement yourself:

• ending the call when the other person says "goodbye". The model won't hang up on its own;
• detecting an answering machine and deciding what to do with it;
• playing a recorded greeting word for word, for example for legal reasons;
• tracking interruptions, when a person starts talking over the agent.

These are all familiar pain points of voice agents. They are exactly what specialized platforms address. So it makes sense to think of the new feature as a transport, not as a ready-made contact center.

Languages and voice quality

The list has 21 voices, but the author found voice acting among them only for English, Portuguese, German, French, Japanese, and Korean. Russian works, but with an accent. For a Russian-language product this is no small matter: on the phone, people quickly notice unnatural speech and hang up more often. If your scenario is not in English, first test the voice on live calls, not on a demo.

The author's overall impression: OpenAI has good voices, but the release looks like a beta.

What to do about it

If you're building a voice agent for appointment booking, support, or lead qualification, it's worth building a SIP prototype and comparing it with your current stack on three parameters: latency, the total cost per minute including the model, and voice quality in your language. If you're currently paying an intermediary for simple call routing, you now have a reason to revisit the contract.

On the other hand, if you need answering machine detection, rigid scripts, and careful handling of interruptions, ready-made platforms are more convenient for now. The news is also worrying for voice platform providers: part of their value is flowing into OpenAI's own API. But for now this is strictly a beta, and it's worth measuring on your own calls.