It looks like they’ve just released the old model as GA:
gemini-live-2.5-flash-native-audio |
|---|
gemini-live-2.5-flash-native-audio has the same issue. It requires audio output to work. I’m actually not able to get it to do anything other than throw errors at the moment, sigh.
I think it’s safe to say Google isn’t planning to support text output with the Live API in the future. If you rely on this functionality, I would suggest investing the time now to adjust/migrate/etc.
thank you very much. this just broke my prod
Yeah, I saw that and got excited, but it’s the native-audio model, which is the broken one.
I will admit that I tested the new Gemini 3 flash non-realtime model (via the REST APIs), and it has really impressive latency. But when it comes to function calling, where the model has to go back and forth with my service, it’s still 4-5x slower than the live API.
The other big problem is that the REST APIs don’t support enabling Google Search Grounding and defining custom tools; you can only do one or the other. The live API allows you to enable Google tools and add custom tools. Given that I assume the Google tools are basically appending the Google tool in the tool list, I have no idea why the REST APIs don’t allow it, but they don’t. 
I still have an outstanding support ticket within the Google Cloud console. The support person has been really helpful and responsive. He has confirmed the issue we are facing and that the model page claims text-to-text should work. He sent it to the engineering team to find out what’s wrong, but I haven’t received any updates in about a week.
This detailed review of the gemini 3 flash attempt has saved me hours. I am new to openai realtime though, and better documentation that the docs? It feels a lot
+1
I have the same problem…
Just a couple of months ago, Google stated in their documentation that [“It [half-cascaded architecture] offers better performance and reliability in production environments, especially with tool use.“]20251025221058/google.dev/gemini-api/docs/live) when comparing the half-cascaded audio with native audio. Maybe this has changed, but after reading this thread, it doesn’t seem like it.
Why deprecate the model when the newer one has no support for text output? Was this an error or is there some kind of move from Google here?
I believe the issue is capture correctly here gemini-live-2.5-flash-preview doesn't call tools · Issue #1698 · google/adk-python · GitHub we need the exp model carried forward into a 3.0 model. @Lalit_Kumar can you provide some guidance on a live 3.0 model that is text only that supports both generateContent API and bidiGenerateContent API. We don’t want to be forced to use the Audio modality if we are spinning our own TTS model.
There aren’t currently any Gemini 3 models that support the live API at this time. Right now, there is only one model at all that supports the live API, and it’s the 2.5 flash voice model. It does not support text in and text out, which is the issue, even though the model website says it does.
I still have a support ticket open with the Google Cloud support team. They reached out to the engineering team to confirm whether text-to-text is supported. But we haven’t heard anything back yet, I’m hoping we do after the holidays.
At this time, you have to use either the Gemini 3 flash REST API or another service, like OpenAI’s live API.
Hey man , did you receive any response from the support team regarding the native-audio not allowing text output ?
Yeah, I have a less-than-ideal conclusion to my quest. The support agent was very helpful, very kind, and very on top of it. I really appreciated him. He was able to get me into a private GA where the model is still available. He suggested others in this thread can also ask their Google Cloud account representative, who might be able to get them access to the private GA.
But beyond that, either we didn’t get anything back from engineering, or they didn’t want to comment. They acknowledged that the model website doesn’t match what the actual model is doing, but they don’t seem interested in resolving it.
The last thing said was to “keep a watch out for future model releases and updates,” which may be a hint… or just something they always say. Who knows.
Hey All!
Has there been any fixes? I really want to use this.