For some reason, when my system tries to create a websocket connection with the model ID “gemini-live-2.5-flash,” the websocket connection is failing with error code 1008. It was working perfectly yesterday and the days before, but now the “gemini-live-2.5-flash” model doesn’t work. I tried setting the model to “gemini-live-2.5-flash-preview-native-audio” and the websocket sets up correctly, but the model doesn’t have the same function calling ability as the “gemini-live-2.5-flash” model.
Hey Alex, did you find a solution for this? I got the exact same error on the same day while it was working for more than a week, and no matter how many times I tried, I couldn’t get gemini-live-2.5-flash back up. I ended up switching to **gemini-2.0-flash-live-preview-04-09**, which technically works, but let’s be real—it’s nowhere near as good, its just awful in regards to latency and following system instructions. gemini-live-2.5-flash was the backbone of my project. I’m also running it through Vertex AI since I need the Gemini Live API to hit the RAG Engine for tool calls. If you’ve managed to find a workaround or fix, I’d love to hear it.
This is exactly the same thing I did. Gemini-2.0-flash-live-preview-04-09 is nowhere near as good and constantly “slips” when trying to call functions. It’s less reliable, and all of my attempts to contact support have failed with no avail.
Gemini-live-2.5-flash was the backbone of my project too, and this has ground development to a halt.
I’m having the same problem. It stopped working at around 4pm MST on Thursday for me. I’m in some kind of workshop with the Google for Cloud Startups program and I’ve emailed my contacts about it. I was using the same model: gemini-live-2.5-flash with a Vertex AI endpoint inside of a websocket. I checked vertex AI “stream realtime” in the google cloud console and the model had disappeared from there also. Maybe they found some kind of bug? Or maybe they gated it? Looking at the Live API docs, it shows Private GA for gemini-live-2.5-flash even though that is for the Gen AI SDK:
I have also defaulted to Gemini-2.0-flash-live-preview-04-09.
I can let you know if I hear anything from my contacts. Hopefully, we will be able to use this model again. It was so much better than this preview model and I would really like to use it in produciton.
It seems the Vertex team may not be able to assist us further. In the meantime, I tried a different approach by using gemini-live-2.5-flash-preview directly through the Gemini Developer API available on Google AI Studio. The documentation, along with sample code, can be found here: https://ai.google.dev/gemini-api/docs/live.
This method worked for me, and I believe it should work for you as well.
Thanks, Usman. Looks like this workaround will help most people. I just wanted to add a thought. It’s possible the model was pulled for legal/compliance reasons. Generally, any non-preview model in Vertex AI is supposed to be HIPAA-compliant under a signed BAA, so if something slipped through early they might have had to retract it. Obviously pure speculation, but it could explain why support hasn’t been able to give us a clear answer.
You’re welcome. I actually thought the same at first, but here’s what I found. I reached out to Vertex AI support a couple of times about this issue, and their response was that you need to submit a request through the support page in your Google Cloud Console to get approval for using that model in your project. The catch is that this goes through Google’s paid support. For individuals like us who don’t have a support subscription, the only option is to rely on the documentation, which isn’t really helpful in this case. So, in short—if you have a support plan, you can raise a request and they’ll likely approve access to that model in Vertex AI.
I get the 1008 error with the model “gemini-2.5-flash-native-audio-preview-09-2025” on Vertex AI. This model works on AI Studio. How to make it work on Vertex AI ? It has another name on Vertex AI ?