Don't wait: your Gemini TTS preview can fail November 17. Google's deprecation page lists November 17, 2026 as the earliest shutdown date for five audio preview models:

Model ID Move to
gemini-3.1-flash-tts-preview gemini-3.8-flash-tts or gemini-3.8-flash-lite-tts
gemini-2.5-flash-preview-tts same
gemini-2.5-pro-preview-tts same
gemini-3.1-flash-live-preview gemini-3.8-live
gemini-2.5-flash-native-audio-preview-12-2025 gemini-3.8-live

"Earliest" is the important word

A fixed date gives you a deadline. We'd rather have one. An earliest date means the calls can stop on any day from then on, without another announcement. Preview models also tend to get shorter notice than stable ones, because they were never promised to last.

Why audio previews are easy to miss

Voice features are often built fast: a read-aloud button, a voice mode for a demo, a live agent prototype that quietly became a feature. They're rarely the core of the app, so nobody owns the migration until the first bug report. It's the same pattern behind most quiet breakages.

What to do today

  1. Search code and config:
grep -rnE "preview-tts|native-audio-preview|flash-live-preview" . 

Then check environment variables and anywhere a model name is stored as data.

  1. Move TTS calls to gemini-3.8-flash-tts (or gemini-3.8-flash-lite-tts if cost matters more than quality), and Live calls to gemini-3.8-live.
  2. Listen to the output. Voices, pacing and pronunciation change between models. A test that only checks "audio came back" won't catch a voice your users don't recognise.
  3. Run one real Live session end to end. Session settings that worked on a preview may not carry over unchanged.

Keep the next one off your error logs

Versyka tracks shutdown dates from each provider's own page, including the "earliest" ones, and finds the model IDs in your code so the swap can be a reviewed pull request instead of an incident.

Is there a preview model running in production in your app right now?