+1 this is very strange to discover and appreciate the PR!

+1 totally agree, we should have possibility to change the model.

+1 agree!

+1 !!

❤A new model gpt-4o-realtime-preview-2024-12-17 has been released with lower price.❤
Thank you.

please see: #92
#92 This PR introduces the ability to select a custom OpenAI realtime model in the RealtimeClient. Users can now specify the desired model for more flexibility and tailored API usage.

As a workaround, you can call client.realtime.connect directly:

export const DEFAULT_REALTIME_MODEL = "gpt-4o-realtime-preview-2024-12-17";
// ...
await client.realtime.connect({ model: DEFAULT_REALTIME_MODEL });
await client.updateSession();

The connect method in RealtimeClient is just a thin wrapper around RealtimeAPI.connect.

As a workaround, you can call client.realtime.connect directly:

export const DEFAULT_REALTIME_MODEL = "gpt-4o-realtime-preview-2024-12-17";
// ...
await client.realtime.connect({ model: DEFAULT_REALTIME_MODEL });
await client.updateSession();
The connect method in RealtimeClient is just a thin wrapper around RealtimeAPI.connect.

Do you maybe have an idea why function calling does not work when I define model like this instead of using client.connect()

I've tried the method above, even though you can see chatgpt-4o-mini-model in websocket tab, You still end up getting charged for chatgpt 4 main model on your usage history. No usage logs for 4o mini

Odd, it's definitely working for me! Are you using a relay?

Odd, it's definitely working for me! Are you using a relay?

Yup, aren't you getting charged for gpt-4o-realtime-preview instead of getting gpt-4o-mini-realtime-preview?

Odd, it's definitely working for me! Are you using a relay?

Yup, aren't you getting charged for gpt-4o-realtime-preview instead of getting gpt-4o-mini-realtime-preview?

The code I suggested needs to be run on the relay server, not the client.

Yes, but it's gpt-4o-realtime-preview-2024-12-17, which is different than the hardcoded version (gpt-4o-realtime-preview-2024-10-01).

As a workaround, you can call client.realtime.connect directly:
export const DEFAULT_REALTIME_MODEL = "gpt-4o-realtime-preview-2024-12-17";
// ...
await client.realtime.connect({ model: DEFAULT_REALTIME_MODEL });
await client.updateSession();
The connect method in RealtimeClient is just a thin wrapper around RealtimeAPI.connect.

Do you maybe have an idea why function calling does not work when I define model like this instead of using client.connect()

Not sure. I use function calling in my application and noticed no difference when I swapped models. The realtime models' tool-calling abilities have seemed far more sensitive to temperature than I've expected. If you're not already, maybe try using the default value for temperature.

Quick Fix -- Go and edit the hardcoded Node file and move it to gpt-4o-realtime-preview-2024-12-17 app.js

I did more digging around. The relay server issue is coming from this. If you're using a relay server you definitely need the fix below
#52

Just fork the repo and add this fix and link it back to your project. That's how I did it. Check devtools of relay server

It really is this easy: #89 (comment)

You just have to make those changes in the right place (i.e., where you're actually connecting to OpenAI :)