From 3686f6a005cd1ca1915aa2542256ad8470eb35bf Mon Sep 17 00:00:00 2001
From: jibanez-staticduo <staticduo@gmail.com>
Date: Wed, 9 Sep 2026 11:17:09 +0200
Subject: [PATCH] fix(chatgpt): select live transport from model metadata

---
 litellm/llms/chatgpt/realtime.py              | 11 +++++++-
 ...odel_prices_and_context_window_backup.json | 10 ++++++++
 model_prices_and_context_window.json          | 10 ++++++++
 .../llms/chatgpt/test_realtime.py             | 25 ++++++++++++++++++-
 4 files changed, 54 insertions(+), 2 deletions(-)

diff --git a/litellm/llms/chatgpt/realtime.py b/litellm/llms/chatgpt/realtime.py
index 66a703eb616..93a79b7d159 100644
--- a/litellm/llms/chatgpt/realtime.py
+++ b/litellm/llms/chatgpt/realtime.py
@@ -9,6 +9,7 @@ from litellm.llms.openai.realtime.handler import OpenAIRealtime
 from litellm.llms.openai.realtime.http_transformation import OpenAIRealtimeHTTPConfig
 from litellm.types.realtime import RealtimeQueryParams
 from litellm.types.router import GenericLiteLLMParams
+from litellm.utils import get_model_info
 
 from .common_utils import CHATGPT_API_BASE
 from .responses.transformation import ChatGPTResponsesAPIConfig
@@ -34,6 +35,14 @@ def realtime_headers(
     }
 
 
+def realtime_endpoint(model: str) -> str:
+    try:
+        model_info: Final = get_model_info(model, custom_llm_provider="chatgpt")
+    except Exception:
+        return "realtime"
+    return "live" if "/v1/live" in (model_info.get("supported_endpoints") or ()) else "realtime"
+
+
 class ChatGPTRealtime(OpenAIRealtime):
     def __init__(self, params: GenericLiteLLMParams, headers: Mapping[str, str]) -> None:
         super().__init__()
@@ -50,7 +59,7 @@ class ChatGPTRealtime(OpenAIRealtime):
 
     def _construct_url(self, api_base: str, query_params: RealtimeQueryParams) -> str:
         base: Final = URL(api_base)
-        endpoint: Final = "live" if query_params.get("model") == "gpt-live-1-codex" else "realtime"
+        endpoint: Final = realtime_endpoint(query_params.get("model", ""))
         if self._call_id:
             return str(
                 base.copy_with(
diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json
index 54ebdc85be9..ae4432e167d 100644
--- a/litellm/model_prices_and_context_window_backup.json
+++ b/litellm/model_prices_and_context_window_backup.json
@@ -27832,6 +27832,16 @@
         "max_tokens": 8191,
         "mode": "embedding"
     },
+    "chatgpt/gpt-live-1-codex": {
+        "litellm_provider": "chatgpt",
+        "mode": "realtime",
+        "supported_endpoints": [
+            "/v1/realtime/calls",
+            "/v1/live"
+        ],
+        "supports_audio_input": true,
+        "supports_audio_output": true
+    },
     "chatgpt/gpt-5.4": {
         "litellm_provider": "chatgpt",
         "max_input_tokens": 1050000,
diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json
index 54ebdc85be9..ae4432e167d 100644
--- a/model_prices_and_context_window.json
+++ b/model_prices_and_context_window.json
@@ -27832,6 +27832,16 @@
         "max_tokens": 8191,
         "mode": "embedding"
     },
+    "chatgpt/gpt-live-1-codex": {
+        "litellm_provider": "chatgpt",
+        "mode": "realtime",
+        "supported_endpoints": [
+            "/v1/realtime/calls",
+            "/v1/live"
+        ],
+        "supports_audio_input": true,
+        "supports_audio_output": true
+    },
     "chatgpt/gpt-5.4": {
         "litellm_provider": "chatgpt",
         "max_input_tokens": 1050000,
diff --git a/tests/test_litellm/llms/chatgpt/test_realtime.py b/tests/test_litellm/llms/chatgpt/test_realtime.py
index 7ce8d5f298a..239a5a12280 100644
--- a/tests/test_litellm/llms/chatgpt/test_realtime.py
+++ b/tests/test_litellm/llms/chatgpt/test_realtime.py
@@ -40,7 +40,7 @@ async def test_chatgpt_call_keeps_oauth_and_frameless_session(chatgpt_tokens):
 
 
 @pytest.mark.parametrize("model,endpoint", [("gpt-realtime-1.5", "realtime"), ("gpt-live-1-codex", "live")])
-def test_realtime_uses_platform_endpoint_with_oauth_headers(model, endpoint, chatgpt_tokens):
+def test_realtime_uses_platform_endpoint_with_oauth_headers(model, endpoint, chatgpt_tokens, local_model_cost_map):
     handler = ChatGPTRealtime(
         GenericLiteLLMParams(),
         {
@@ -55,3 +55,26 @@ def test_realtime_uses_platform_endpoint_with_oauth_headers(model, endpoint, cha
     assert headers["Authorization"] == "Bearer test-token-default"
     assert "authorization" not in headers
     assert headers["openai-alpha"] == "quicksilver=v2"
+
+
+@pytest.mark.parametrize("endpoint", ["live", "realtime"])
+@pytest.mark.parametrize("call_id", [None, "rtc_metadata"])
+def test_realtime_routes_new_models_using_registered_metadata(endpoint, call_id, chatgpt_tokens, local_model_cost_map):
+    model = "metadata-voice-model"
+    litellm.register_model({f"chatgpt/{model}": {
+        "litellm_provider": "chatgpt", "mode": "realtime", "supported_endpoints": [f"/v1/{endpoint}"]
+    }})
+    handler = ChatGPTRealtime(GenericLiteLLMParams(chatgpt_realtime_call_id=call_id), {})
+    expected = (
+        f"wss://api.openai.com/v1/{endpoint}?model={model}" if call_id is None
+        else f"wss://api.openai.com/v1/live/{call_id}" if endpoint == "live"
+        else f"wss://api.openai.com/v1/realtime?call_id={call_id}"
+    )
+    assert handler._construct_url("https://api.openai.com/v1", {"model": model}) == expected
+
+
+def test_realtime_unknown_model_keeps_standard_endpoint(chatgpt_tokens, local_model_cost_map):
+    handler = ChatGPTRealtime(GenericLiteLLMParams(), {})
+    assert handler._construct_url("https://api.openai.com/v1", {"model": "unknown-voice-model"}) == (
+        "wss://api.openai.com/v1/realtime?model=unknown-voice-model"
+    )