Gemini 3.8 Live Extended Thinking

Google · Audio Generation

WEBSOCKET /v1/realtime

Voice conversation that keeps talking while it reasons in the background, with 30 voices, live video input, tool calling, and adjustable reasoning effort.

At a glance

Pricing

Example request

import asyncio, json, os, websockets
URL = "wss://api.empiriolabs.ai/v1/realtime?model=gemini-3-8-live-extended-thinking"
async def main():
    async with websockets.connect(
        URL,
        additional_headers=[("Authorization", f"Bearer {os.environ['EMPIRIOLABS_API_KEY']}")],
        max_size=None,
    ) as ws:
        print(json.loads(await ws.recv())["type"])  # session.created
        await ws.send(json.dumps({
            "type": "conversation.item.create",
            "item": {"type": "message", "role": "user",
                     "content": [{"type": "input_text", "text": "Say hello."}]},
        }))
        await ws.send(json.dumps({"type": "response.create"}))
        async for raw in ws:
            event = json.loads(raw)
            if event["type"] == "response.done":
                break
asyncio.run(main())

Parameters

Notes

Connecting

Full-duplex voice over a WebSocket at wss://api.empiriolabs.ai/v1/realtime?model=gemini-3-8-live-extended-thinking, authenticated with the ordinary Authorization Bearer header. Configure the session with session.update before you send the first audio: voice, instructions, turn detection and tools. These settings are fixed once the conversation starts.

Sending audio and video

• Append microphone audio as input_audio_buffer.append events: 16-bit PCM at 16 kHz, or at 24 kHz with input_audio_format set to pcm24.
• Append live video frames as input_image_buffer.append events, as base64 JPEG or PNG images.
• Voice activity detection is on by default, so keep streaming for about a second after the speech ends. Set turn_detection to null to end each turn yourself with input_audio_buffer.commit.

Voices and language

• Use one of the 30 voices in the voice list, for example Puck, Kore or Charon. Any other value is refused.
• The model replies in the language you speak.

Replies

• Speech streams as response.audio.delta events, 16-bit PCM at 24 kHz, with the words as response.audio_transcript.delta events. Your own speech is transcribed as well.
• Speaking while the model replies interrupts it.

Reasoning

• reasoning_effort sets how much the model reasons in the background while it talks: low, medium (the default) or high. Set it before the conversation starts.
• The model often acknowledges a question first and answers in a second turn. Each turn is billed on its own.

Tools

• Function calling with JSON Schema parameters. Return each result as a function_call_output item with its call_id.

Billing

Audio and text tokens are priced separately in both directions, video frames bill as input tokens, and reasoning tokens bill as output tokens. Each completed turn is billed on its own from the usage the model reports, including a reply you interrupt.

Machine-readable schema: GET https://api.empiriolabs.ai/v1/models/gemini-3-8-live-extended-thinking.