Skip to content
Nappai Documentation

OpenAI Realtime

OpenAI Realtime is one of the models you can connect to a Voice Agent Realtime. A single OpenAI model hears the caller, decides when they have finished, thinks and answers in its own voice.

It plugs into the Realtime model connector of the Voice Agent Realtime node. It holds the model, voice and credential the agent uses on every call.

With the gpt-realtime models, the caller’s audio goes through OpenAI’s noise reduction for the near voice of a phone or a headset, so the model understands better with background noise. Nothing to configure.

When to use it

  • You already use OpenAI and want the same credential for your voice agent.
  • You want to choose between a full model and a cheaper one for simple calls.
  • You want an agent that keeps listening while it talks, with gpt-live-1.

Configuration

You need an OpenAI API credential with your OpenAI account key. See Credentials.

FieldWhat it does
Credential (required)The OpenAI API credential used to connect to the service.
Modelgpt-realtime-2 (default) uses the tools itself. gpt-realtime-mini is the cheapest. gpt-live-1 keeps listening while it talks and hands tool use to a second model. gpt-realtime is also available. You can type another model name.
VoiceThe voice it speaks with: marin (default), cedar, alloy, ash, ballad, coral, echo, sage, shimmer or verse. With gpt-live-1, aster, beacon, cinder, stone and vesper are also available. All of them speak any language.

Advanced options

FieldWhat it does
Answer EagernessHow soon it decides the caller’s sentence is over: auto (default), low, medium or high. high answers sooner but may cut people off; low waits longer. Not used by gpt-live-1, which takes turns by itself.
GPT-Live Tools ModelOnly with gpt-live-1: the model that uses the tools it hands over. Default: gpt-6-luna. Billed separately.

Output

The output holds the model settings. It connects to the Realtime model connector of the Voice Agent Realtime node.

Example

An online shop wants an agent that answers quickly about order status.

  1. Add OpenAI Realtime and connect it to Realtime model on the Voice Agent Realtime.
  2. In Credential, choose the OpenAI API credential.
  3. Leave Model on gpt-realtime-2 and pick a voice in Voice.
  4. Test the agent in the playground and listen to how it sounds.

Good practices

  • Try several voices in the playground before deciding: all of them speak your language, but each has its own tone.
  • The Greeting and fixed farewells may come out with a word or two changed: the model says them its own way.
  • Response time varies. On a test call on our development server on 14 October 2026, gpt-realtime-2 took between 1.5 and 1.9 seconds from the moment the caller stopped speaking to the first sound of its answer. Check it in the playground with your own calls.
  • OpenAI bills the usage to your own OpenAI account. Check the current price there.