Two voice models
Every agent answers calls with one of two voice models:- Classic is the voice model agents have always used. Every agent starts on Classic, and nothing changes until you pick another model.
- GPT-Live is OpenAI’s live voice model. It listens and talks at the same time, so conversations flow quickly and callers can interrupt naturally.
Choose a model
- Open Agents and select your agent.
- In the identity panel on the left, open Voice model and pick Classic or GPT-Live.
- The change saves straight away. The next call uses the new model.
Choose a voice
Under Voice, pick who answers your calls: Ava, Ivy, Henry, Jess, Sophia, Angel, Oscar, Alex or Sid. Each one has its own voice and manner, and a voice on each model, so your pick works on Classic and GPT-Live.- Hear greeting plays your agent’s own greeting in the voice callers will hear on the model you picked.
- Test call lets you talk to your agent from the browser. Nothing is dialled, test calls last up to 3 minutes, and they are not billed. They appear in your call log.
What each model is good at
Pick GPT-Live when you want calls that feel quick and natural, and your callers ring from quiet places. Stay on Classic when callers often name hard-to-hear suburbs or streets, when your callers need extra time to answer, or when you want a specific ElevenLabs voice.
Limits of GPT-Live
These come from OpenAI’s platform. They apply to every agent on GPT-Live.Prompt length
GPT-Live accepts about 16,000 tokens of prompt, roughly 12,000 words, including your listings.- If your prompt is longer, the end of it is left out of what GPT-Live reads before each call. Your agent still uses the full prompt when it looks up an answer.
- If more than a fifth of your prompt would be left out, GPT-Live shows as unavailable with the reason “This agent’s prompt is too long for GPT-Live”, and the agent stays on Classic. Shorten the prompt to use GPT-Live.
Its own voices
GPT-Live makes its own speech and can’t use ElevenLabs voices. Each character’s GPT-Live voice is made to sound close to its Classic voice, but they are not identical. On a small share of calls, about 1 in 100, GPT-Live uses a standard voice instead. OpenAI asks that callers are told the voice they hear is AI-generated, so on GPT-Live your agent adds “I’m an AI assistant.” to its greeting, just before the greeting’s closing question, or at the end if there is none. For example, “Hi, you’ve called Harbour & Co, Ava speaking. How can I help?” becomes “Hi, you’ve called Harbour & Co, Ava speaking. I’m an AI assistant. How can I help?” If your greeting already says so, for example “Ava, Harbour & Co’s AI assistant”, it is kept exactly as you wrote it. Hear greeting plays the greeting as callers hear it. Classic greetings are not changed.Place names
GPT-Live has no setting that helps it hear particular words. Your Specialised Terms are given to the agent as background, which helps, but suburb and street names are mostly heard by sound.Turn-taking
GPT-Live decides by itself when the caller has finished. Patience has no effect on GPT-Live, and there is no setting for background noise. Speech in the background, such as a TV or another person, is sometimes answered as if the caller said it.Where calls are processed
GPT-Live processes calls outside Australia, in the United States. OpenAI has no Australian region for it.When Classic answers instead
GPT-Live never leaves a caller waiting. Classic answers the call when:- GPT-Live is unavailable for the agent (the reason shows under Voice model);
- GPT-Live doesn’t start in time, or OpenAI is busy;
- GPT-Live is switched off for a short time while we fix a problem. The list then shows “GPT-Live is switched off for now”.