Live API: response_modalities=[TEXT] causes 1011 Internal Error on gemini-3.1-flash-live-preview

[sniper-fly]

Environment details

• Programming language: Python
• OS: Linux (Ubuntu)
• Language runtime version: 3.13.2
• Package version: 1.70.0

Description

When connecting to gemini-3.1-flash-live-preview via the Live API with response_modalities=[TEXT], the connection immediately fails with a 1011 Internal error encountered WebSocket close frame.

Using response_modalities=[AUDIO] with the same model and API key works without issues. The workaround is to use AUDIO modality with output_audio_transcription to obtain text responses, but this is suboptimal for text-only use cases.

Steps to reproduce

import asyncio
from google import genai
from google.genai import types

client = genai.Client(api_key="YOUR_API_KEY")

async def main():
    # This FAILS with 1011
    config_text = types.LiveConnectConfig(
        response_modalities=[types.Modality.TEXT],
    )
    async with client.aio.live.connect(
        model="gemini-3.1-flash-live-preview", config=config_text
    ) as session:
        await session.send_realtime_input(text="Hello")
        async for msg in session.receive():
            print(msg.text)

asyncio.run(main())

Error

websockets.exceptions.ConnectionClosedError: received 1011 (internal error) Internal error encountered.

google.genai.errors.APIError: 1011 None. Internal error encountered.

Working comparison (AUDIO modality)

# This WORKS
config_audio = types.LiveConnectConfig(
    response_modalities=[types.Modality.AUDIO],
    speech_config=types.SpeechConfig(
        voice_config=types.VoiceConfig(
            prebuilt_voice_config=types.PrebuiltVoiceConfig(voice_name="Puck")
        )
    ),
    output_audio_transcription=types.AudioTranscriptionConfig(),
)
async with client.aio.live.connect(
    model="gemini-3.1-flash-live-preview", config=config_audio
) as session:
    # Connected successfully, receives audio + transcription
    ...

Additional notes

• Tested with both v1alpha and v1beta API versions — same result
• Tested with raw WebSocket (bypassing the SDK) — same result
• Tested with ephemeral tokens + BidiGenerateContentConstrained — same result
• The model is listed in client.models.list() and is the only model supporting bidiGenerateContent
• Regular (non-Live) API calls work fine with the same API key