GPT-Live now runs on your LiveKit credentials

LiveKit Agents 1.8.8 lets you run OpenAI's GPT-Live with your LiveKit credentials, so a voice agent no longer needs its own OpenAI key for it. Pass the model string as the session's LLM, and LiveKit Inference serves it through a new class called DuplexModel.

Slide text

livekit-agents 1.8.8

GPT-Live now runs on your LiveKit credentials

session = AgentSession(llm="openai/gpt-live-1")

No separate OpenAI key. The string resolves to inference.DuplexModel.

STT → LLM → TTS REALTIME DECISION

GPT-Live speaks with the user and hands delegated work to a backend model. Through Inference that backend defaults to GPT 5.6 Luna, and your function tools work as usual. There are limits. Delegated work runs on the default service tier only, the backend has to be an OpenAI model that Inference serves, and hosted tools like web search are refused with an error. For those, use the OpenAI plugin directly.

Slide text

inference/realtime/gpt_live.py

Delegated work runs on a backend model that Inference serves

AgentSession

llm="openai/gpt-live-1"

GPT-Live

default voice "marin"

delegation

Backend model

gpt-5.6-luna by default

Limits through Inference

Default service tier only, for delegated work

The backend must be an OpenAI LLM that Inference serves

Hosted tools such as WebSearch() are refused; function tools work

STT → LLM → TTS REALTIME DECISION

Other realtime models get the same one-string setup. Inference now accepts three Gemini Live model strings, including an extended-thinking variant. Grok Voice 3.0 replaces Think Fast 2.0 in both the Inference list and the xAI plugin. And GPT-Live 1 is the one full-duplex model on the list.

Slide text

inference/_realtime_models.py

Gemini Live and Grok Voice 3.0 work as model strings too

Google

Gemini Live

google/gemini-3.8-live

google/gemini-3.8-live-
extended-thinking

google/gemini-live

xAI

Grok Voice 3.0

xai/grok-voice-3.0

xai/grok-voice-
think-fast-2.0

OpenAI

GPT-Live 1

openai/gpt-live-1

STT → LLM → TTS REALTIME DECISION

There's a new kind of model in the framework. A DecisionModel takes the conversation and a batch of named questions, and answers them all in one request, without delaying the agent's reply. The first provider is TypeSafe's Jev, through its own API or OpenRouter's Decisions API. Each question is a probability, a choice between named options, or a score on levels you describe, like calm, annoyed or angry.

Slide text

livekit-plugins-typesafe

Decision models answer questions about a call in the background

model = typesafe.DecisionModel(model="jev-latest")
response = await model.evaluate(
    input=context,
    decisions={
        "handoff": llm.Probability(
            "The caller requests a human."),
        "intent": llm.Choice("What does the caller want?",
            options={"booking": …, "other": …}),
        "frustration": llm.Score("How frustrated…?",
            levels=["Calm", "Annoyed", "Angry"]),
    },
)

Probability

0 to 1 that a statement is true

Choice

one named option

Score

a position on ordered levels

STT → LLM → TTS REALTIME DECISION

Two changes can break existing code. The Inference RealtimeModel no longer subclasses the OpenAI plugin's RealtimeModel, so a type check against the plugin class stops matching Inference models. Check for both classes. Incidental wildcard exports, like the realtime model module's logger, are gone too. And for GPT-Live, the unsupported aster voice is replaced by cedar, so update any voice set to aster.

Slide text

Breaking changes

Two changes can break an existing agent

1.8.7

isinstance(m,
    openai.realtime.RealtimeModel)
# True for inference.RealtimeModel

1.8.8

isinstance(m, (
    openai.realtime.RealtimeModel,
    inference.RealtimeModel))

1.8.7

GPTLiveModel(voice="aster")

1.8.8

GPTLiveModel(voice="cedar")

STT → LLM → TTS REALTIME DECISION

Retired GPT Realtime and Grok Voice IDs keep working. Each one logs a deprecation warning that names its replacement, so you can move over when it suits you. The realtime inference class argument is deprecated as well. It's no longer sent and it's due to be removed in version two.

Slide text

Behavior changes

Retired model IDs still connect, with a deprecation warning

openai/gpt-realtime-2	→	openai/gpt-realtime-2.1
openai/gpt-realtime-1.5	→	openai/gpt-realtime-2.1
xai/grok-voice-think-fast-2.0	→	xai/grok-voice-3.0
RealtimeModel(inference_class=)	→	no effect, removed in v2.0

STT → LLM → TTS REALTIME DECISION

If you call generate reply with instructions on Gemini 3.1 or 3.8, those instructions now go in as a user turn. Sent as a model turn, they made the model speak its reasoning or stay silent, and the wrapper around the new user turn keeps the model from quoting them back later as the user's words. Hosted xAI sessions now emit one final user transcript per turn. And in TTS chunking, a stray closing tag can no longer cancel out a later opening one.

Slide text

Fixes

Gemini 3.x reply instructions now go in as a user turn

Gemini 3.x

Sent as a model turn, they made the model speak its reasoning or stay silent. The wrapper keeps it from quoting them back as the user's words.

xAI

Hosted sessions emit one final user transcript per turn.

TTS chunking

A stray closing XML tag no longer cancels out a later opening tag.

STT → LLM → TTS REALTIME DECISION

Before you upgrade, look for two things: type checks against the OpenAI plugin's RealtimeModel, and GPT-Live agents using the aster voice. Fix those first. Then install 1.8.8 with the extras you already use, and add the TypeSafe plugin if you want decision models. GPT-Live now runs on your LiveKit credentials.

Slide text

Upgrade

Upgrade to 1.8.8 after two quick checks

Type checks against openai.realtime.RealtimeModel

GPT-Live agents set to the aster voice

$ pip install -U
  "livekit-agents[openai]==1.8.8"

$ pip install livekit-plugins-typesafe

STT → LLM → TTS REALTIME DECISION