OpenAI Live

Experimental Feature

This feature is experimental and may change in future versions. Use with caution in production environments.

Language Support

This feature is only supported in python.

OpenAI GPT-Live brings continuous voice conversations to your agents. The Live model listens and speaks while a separate Responses model handles reasoning and selects tools for BidiAgent to execute.

Follow the quickstart to build a voice agent, using the Live configuration below.

Installation

Install the OpenAI Live extra in a separate virtual environment:

pip install "strands-agents[bidi-openai-live]"

For local audio dependencies and device setup, see Audio I/O.

Credentials

Create an OpenAI API key with access to your chosen Live and Responses models, then set it in your environment:

export OPENAI_API_KEY="your-api-key"

OpenAILiveModel reads OPENAI_API_KEY automatically. Pass OpenAI client options through client_args, such as client_args={"api_key": "your-api-key"}.

Configuration

Configure a Live model for speech and a Responses model for reasoning and tool use, then pass the provider to BidiAgent:

from strands.bidi.agent import BidiAgent
from strands.bidi.models import OpenAILiveModel

model = OpenAILiveModel(
    model_id="gpt-live-1",
    delegation_config={"type": "responses", "model_id": "gpt-6-luna"},
    voice="marin",
    audio_config={"sample_rate": 24000},
)
agent = BidiAgent(model=model)

Both model IDs are required. The top-level model_id selects the voice model; delegation_config["model_id"] selects the Responses model. Only Responses delegation is supported.

Choose an OpenAI voice or omit voice to use the API default. Input and output use mono PCM16 at 24 kHz; set audio_config={"sample_rate": 16000} for 16 kHz. Custom inputs must stream continuously, including silence.

See the OpenAILiveModel API reference for all constructor options.

Delegation settings

Configure backend reasoning and output through delegation_config:

from strands.bidi.models import OpenAILiveModel

model = OpenAILiveModel(
    model_id="gpt-live-1",
    delegation_config={
        "type": "responses",
        "model_id": "gpt-6-luna",
        "max_output_tokens": 512,
        "reasoning": {"effort": "low"},
    },
)

Set delegation_config["system_prompt"] to override the agent’s system prompt for backend reasoning and tool use. Otherwise, both models use the agent’s prompt. Register tools on BidiAgent.

For text settings, service tier, and other options, see OpenAILiveResponsesDelegationConfig and OpenAI’s Responses delegation guide.

Connection restarts

The provider saves sessions with OpenAI and forks them on restart, preserving the conversation and configuration. Forking requires enabled project storage and a finalized recording; it is unavailable with Zero Data Retention.

Automatic restart is disabled

Automatic restart defaults to disabled because Live cannot identify completed spoken turns. Set connection={"auto_restart": True} to enable it, accepting that a restart may cut off speech. See Connection restarts for scheduling options.

Compatibility

When consuming Strands events, account for these differences:

Response stops do not mark completed speech

Live has no event marking the end of each spoken response. The provider creates Strands response boundaries at new backend text streams or tool-use groups. Without such a transition, audio and response streams can remain open until the session closes.

BidiResponseStopEvent does not indicate that speech has finished. Track playback in your audio player, and do not end a conversation on the first response stop.

• Transcripts: Each API fragment becomes a completed transcript block and history message, even when it contains only part of a sentence. Backend text and spoken transcripts both enter history.
• Barge-in: The provider does not emit BidiBargeInEvent. Strands cannot automatically clear buffered audio or cancel tools when the user speaks.
• Usage: Token usage covers the Responses model only. Live reports voice duration in seconds, which Strands token metrics do not include. See OpenAI’s usage guidance.

Tags

bidi-streaming15