litellm - 💡(How to fix) Fix [Feature]: Support gpt-realtime-whisper model (OpenAI Realtime Transcription)

Feature Request

Is your feature request related to a problem? Please describe.

LiteLLM does not currently support the gpt-realtime-whisper model released by OpenAI.

This model is a dedicated real-time speech-to-text model that streams transcription results via the /v1/realtime WebSocket API using the new /v1/realtime/transcription_sessions endpoint. It is distinct from the standard gpt-realtime-* conversation models in that it focuses solely on transcription rather than bidirectional audio conversation.

When attempting to use this model via LiteLLM:

import litellm

response = litellm.speech_to_text(
    model="openai/gpt-realtime-whisper",
    ...
)

It results in an unsupported model error or falls through without proper routing.

Describe the solution you'd like

1. Model registration: Add gpt-realtime-whisper to LiteLLM's OpenAI model list/registry (e.g., in model_prices_and_context_window.json or equivalent).
2. Endpoint routing: Route requests for gpt-realtime-whisper to the /v1/realtime/transcription_sessions WebSocket endpoint, not the standard /v1/realtime sessions endpoint.
3. Pricing: The model is priced at $0.017 / minute (audio input), not per token. Ensure cost tracking reflects this.

Model details

Describe alternatives you've considered

Currently, the only workaround is to call the OpenAI API directly, bypassing LiteLLM entirely, which defeats the purpose of using LiteLLM as a unified proxy/gateway.

Additional context

• OpenAI Realtime Transcription API docs
• Related existing support: gpt-realtime-mini, gpt-4o-realtime-preview (conversation models) are already supported, but the transcription-specific variant is missing.