I am writing to seek urgent clarification regarding the upcoming deprecation and model migration, as we are less than two weeks away from the 2025-12-09 deadline.
- Model Identification & Deprecation
Currently, my production environment relies on gemini-2.5-flash-live-preview, which has been very stable.
The documentation states that gemini-live-2.5-flash-preview will be deprecated on Dec 9, 2025.
Question: Can you confirm if gemini-2.5-flash-live-preview is indeed the model being deprecated?
2. Stability Issues with the Replacement Model
I attempted to migrate to the recommended newer model: gemini-2.5-flash-native-audio-preview-09-2025.
However, I have encountered severe stability issues and regressions compared to the previous version, specifically regarding Function Calling and Instruction Following.
Specific Regressions Observed:
Infinite Tool Calling Loops: The native-audio model frequently gets stuck in a loop, repeatedly calling the same tool/function without processing the result or moving forward. The previous live-preview model handled this correctly.
Ignoring System Instructions: The model often disregards the prompt instructions during the session, which was not an issue with the older model.
The Dilemma
With the Dec 9th deprecation imminent, I am forced to choose between a deprecated model or an unstable one that breaks my application logic.
Is there a more stable version of the native-audio model available?
Will the life of gemini-2.5-flash-live-preview be extended until the regressions in the new model are fixed?
Thanks.
4 Likes
+1 My team is currently experiencing the same dilemma, we’ve been using the gemini-2.5-flash-live-preview with certain stability, while our tests with the native audio model have a lot of latency in the conversation.
As far as we know, the gemini-2.5-flash-live-preview model availability won’t be extended and we don’t really have a solution for this issue.
1 Like
I have exactly the same problem! My use case completely breaks with the native models. They are for me especially bad at prompt instructions.
Please extend the deprecation date 
since gemini-2.5-flash-live-preview is stable as it is
Hi @user3640,
Thanks for reaching out, sorry for the delayed response.
Yes, gemini-live-2.5-flash-preview model was deprecated on Dec 9,2025: https://ai.google.dev/gemini-api/docs/deprecations . The recommended model is gemini-2.5-flash-native-audio-preview-12-2025.
Could you please try using this newer model and let us know if you are still facing any issues along with some details of your configuration and what you are trying to achieve?
Thanks again!
Hi @Srikanta_K_N,
Thank you for your response and the clarification on the model naming.
I have migrated to gemini-2.5-flash-native-audio-preview-12-2025 as recommended, but I’m still experiencing significant stability issues that are impacting my production environment:
Issues Encountered:
- Unstable Tool/Function Calling
The model exhibits inconsistent behavior when calling tools/functions:
Sometimes it gets stuck in infinite loops, repeatedly calling the same function
Other times it fails to call the necessary functions altogether
The previous gemini-2.5-flash-live-preview model handled function calling much more reliably
- Audio-Text Synchronization Issues
When using sendClientContent({ turns: message, turnComplete: true }) to send an initial greeting in Chinese:
The audio continues playing normally
However, the text/subtitles freeze mid-sentence and stop updating
This creates a disconnect between what the user hears and what they see on screen
Questions:
Are these known issues with the December 2025 native-audio model?
Is there a recommended workaround for the audio-text synchronization problem?
Are there any best practices for ensuring stable function calling with this model?
The combination of these issues is making it difficult to maintain a production-ready application. Any guidance would be greatly appreciated.
Hi @user3640,
Thanks for the follow up!
To help us debug this issue, could you please share the minimal reproducible code and prompt?
The new model gemini 3.1 live resolve almost of this issues!