Repeated 1011 Internal error encountered on gemini-2.5-flash-native-audio-preview-12-2025 mid-session (with session resumption)

Gemini API
,
1

Hi team,

We’re running into a persistent stability issue on the Gemini Live native audio preview model and would love guidance on whether this is known / being worked on, or whether something in our config is triggering it.

Setup

  • Model: gemini-2.5-flash-native-audio-preview-12-2025
  • Client: Python, via pipecat-ai 0.0.104 (GeminiLiveLLMService) — WebSocket transport
  • Voice: Kore
  • Tools configured: google_search + a single custom generate_image function declaration
  • Thinking: thinking_budget=0
  • Input audio transcription: enabled
  • Context window compression: enabled (trigger_tokens=10000)
  • Session resumption: enabled (we pass session_resumption_handle on reconnect)
  • System instruction length: ~850 chars
  • Region: US (requests originate from US-East)

Symptom

The WebSocket is closed by the Gemini server with close code 1011 and body None. Internal error encountered. — often within the first 10–30 seconds of a session, and frequently in rapid succession.

Representative session (today, ~3 min long)

T+0s    Connected to Gemini Live
T+8s    1011 "None. Internal error encountered." — reconnect with session_resumption_handle (success)
T+10s   Connection stable for 10s, failure counter reset
T+20s   1011 again — reconnect (success)
T+24s   1011 again — reconnect (success)

Three internal errors in ~26 seconds, all identical, with no apparent pattern in audio content or user input. Session resumption works (the socket comes back), but the bot audio does not resume on its own — from the user’s perspective the model simply stops speaking mid-sentence and never recovers until they re-open the session.

What we’ve observed

  • Error volume is small in absolute terms but affects real user sessions (confirmed via Sentry + PostHog funnel — ~1% of voice sessions today, all premium users).
  • 100% of the 1011s carry the exact string None. Internal error encountered. — no additional detail in the close frame.
  • The errors are not correlated with specific prompts, tool calls, or user utterances — they appear to fire while the server is generating audio.
  • Removing google_search or generate_image tools does not appear to eliminate the issue (anecdotal).
  • GA non-audio Gemini models over the same network path are stable.

Questions

  1. Is 1011 on gemini-2.5-flash-native-audio-preview-12-2025 a known stability issue you’re tracking? Any ETA on improvements before the model graduates from preview?
  2. Is there a way to get more diagnostic detail in the close frame (the current None. Internal error encountered. makes it impossible for us to correlate with specific sessions on your side)?
  3. When session_resumption_handle successfully reconnects after a 1011, is the expected behavior that the model continues its in-flight response, or does it wait for the next user turn? We’re observing the latter — is that intentional?
  4. Any recommended server-side session ID / request ID we can include in bug reports so you can look up the specific failed session in your logs?

Happy to share more logs, request IDs, or reproduce on demand. Thanks for any pointers.