I am trying to migrate to GPT Realtime according to the following, https://docs.vapi.ai/openai-realtime#remove-transcriber-configuration But I cannot “Remove transcriber configuration.” - There is no option to remove it from the web UI. Deepgram is selected by default. - When I try to update using the API at https://docs.vapi.ai/api-reference/assistants/update, I get the following error: 400 - "{\"message\":[\"transcriber.provider must be one of the following values: assembly-ai, azure, custom-transcriber, deepgram, 11labs, gladia, google, openai, talkscriber, speechmatics, cartesia\"],\"error\":\"Bad Request\",\"statusCode\":400}" https://cdn.discordapp.com/attachments/1422037986928758804/1422037988057284772/2025-09-29_10.47.49.png?ex=68db3794&is=68d9e614&hm=df6098479654fba6dba11d8f5cef355b2abd0ab49158d5ac873f6067cde30bac&

To resolve the error, set the transcriber provider to

"openai"

instead of trying to remove the transcriber field. For example:

json
{
  "transcriber": {
    "provider": "openai"
  }
}

This aligns with the migration guide, as realtime models natively handle speech and expect

"openai"

as the provider in the configuration API. See the [Vapi migration guide](https://docs.vapi.ai/openai-realtime#remove-transcriber-configuration) for reference. Source: - [OpenAI Realtime Documentation](https://docs.vapi.ai/openai-realtime)

Is this bot’s answer correct?

We suggest not using realtime since our assistants are configured to use non-thinking models

Sorry, just to clarify — since we are prioritizing low latency in conversations, we’ve been using lightweight models like 4o-mini or 4.1-mini instead of the thinking models. Does using a Realtime model still have any impact in that case?

Even with the lighter models we still recommend avoiding realtime for your use case

Thank you. Why do you recommend not using real-time?

With the way that our platform works and how it transfers data to and from all of the providers, thinking models are expensive in terms of token usage and request calls. For example, if you have a 250 maxToken value for your assistant and the model needs to think, the thinking could take up all 250 tokens and not have enough to respond back or execute a tool call. You would need to increase your maxToken value to 1000-2000 to account for the thinking and the response.