fix: default realtime websocket model to gpt-realtime-1.5 (#2737)

Author: seratch

examples/realtime/twilio_sip/server.py

@@ -69,7 +69,7 @@ async def accept_call(call_id: str) -> None:
             f"/realtime/calls/{call_id}/accept",
             body={
                 "type": "realtime",
-                "model": "gpt-realtime",
+                "model": "gpt-realtime-1.5",
                 "instructions": instructions_payload,
             },
             cast_to=dict,

src/agents/realtime/openai_realtime.py

@@ -158,6 +158,7 @@
 
 
 _USER_AGENT = f"Agents/Python {__version__}"
+DEFAULT_REALTIME_MODEL = "gpt-realtime-1.5"
 
 DEFAULT_MODEL_SETTINGS: RealtimeSessionModelSettings = {
     "voice": "ash",
@@ -271,7 +272,7 @@ class OpenAIRealtimeWebSocketModel(RealtimeModel):
     """A model that uses OpenAI's WebSocket API."""
 
     def __init__(self, *, transport_config: TransportConfig | None = None) -> None:
-        self.model = "gpt-realtime"  # Default model
+        self.model = DEFAULT_REALTIME_MODEL
         self._websocket: ClientConnection | None = None
         self._websocket_task: asyncio.Task[None] | None = None
         self._listeners: list[RealtimeModelListener] = []
@@ -1105,7 +1106,7 @@ def _get_session_config(
         # Construct full session object. `type` will be excluded at serialization time for updates.
         session_create_request = OpenAISessionCreateRequest(
             type="realtime",
-            model=(model_settings.get("model_name") or self.model) or "gpt-realtime",
+            model=(model_settings.get("model_name") or self.model) or DEFAULT_REALTIME_MODEL,
             output_modalities=output_modalities,
             audio=OpenAIRealtimeAudioConfig(
                 input=OpenAIRealtimeAudioInput(**audio_input_args),

tests/realtime/test_conversion_helpers.py

@@ -33,7 +33,7 @@ def test_try_convert_raw_message_valid_session_update(self):
                 "type": "session.update",
                 "other_data": {
                     "session": {
-                        "model": "gpt-realtime",
+                        "model": "gpt-realtime-1.5",
                         "type": "realtime",
                         "modalities": ["text", "audio"],
                         "voice": "ash",

tests/realtime/test_openai_realtime.py

@@ -113,6 +113,36 @@ def mock_create_task_func(coro):
         assert model._websocket_task is not None
         assert model.model == "gpt-4o-realtime-preview"
 
+    @pytest.mark.asyncio
+    async def test_connect_defaults_to_gpt_realtime_1_5(self, model, mock_websocket):
+        """Test that connect() uses gpt-realtime-1.5 when no model is provided."""
+        config = {
+            "api_key": "test-api-key-123",
+            "initial_model_settings": {},
+        }
+
+        async def async_websocket(*args, **kwargs):
+            return mock_websocket
+
+        with patch("websockets.connect", side_effect=async_websocket) as mock_connect:
+            with patch("asyncio.create_task") as mock_create_task:
+                mock_task = AsyncMock()
+
+                def mock_create_task_func(coro):
+                    coro.close()
+                    return mock_task
+
+                mock_create_task.side_effect = mock_create_task_func
+
+                await model.connect(config)
+
+                mock_connect.assert_called_once()
+                call_args = mock_connect.call_args
+                assert call_args[0][0] == "wss://api.openai.com/v1/realtime?model=gpt-realtime-1.5"
+                assert model.model == "gpt-realtime-1.5"
+
+        assert model._websocket_task is not None
+
     @pytest.mark.asyncio
     async def test_session_update_includes_noise_reduction(self, model, mock_websocket):
         """Session.update should pass through input_audio_noise_reduction config."""
@@ -788,6 +818,7 @@ def test_get_and_update_session_config(self, model):
     def test_session_config_defaults_audio_formats_when_not_call(self, model):
         settings: dict[str, Any] = {}
         cfg = model._get_session_config(settings)
+        assert cfg.model == "gpt-realtime-1.5"
         assert cfg.audio is not None
         assert cfg.audio.input is not None
         assert cfg.audio.input.format is not None

tests/realtime/test_realtime_model_settings.py

@@ -51,7 +51,7 @@ def helper() -> str:
 
     monkeypatch.setattr(agent, "get_all_tools", AsyncMock(return_value=[helper]))
     agent.handoffs = [RealtimeAgent(name="handoff-child")]
-    base_settings: RealtimeSessionModelSettings = {"model_name": "gpt-realtime"}
+    base_settings: RealtimeSessionModelSettings = {"model_name": "gpt-realtime-1.5"}
     starting_settings: RealtimeSessionModelSettings = {"voice": "verse"}
     run_config: RealtimeRunConfig = {"tracing_disabled": True}
 
@@ -68,9 +68,9 @@ def helper() -> str:
     assert merged["tools"][0].name == helper.name
     assert merged["handoffs"][0].agent_name == "handoff-child"
     assert merged["voice"] == "verse"
-    assert merged["model_name"] == "gpt-realtime"
+    assert merged["model_name"] == "gpt-realtime-1.5"
     assert merged["tracing"] is None
-    assert base_settings == {"model_name": "gpt-realtime"}
+    assert base_settings == {"model_name": "gpt-realtime-1.5"}
 
 
 @pytest.mark.asyncio

tests/realtime/test_session_payload_and_formats.py

@@ -26,10 +26,10 @@ class _DummyModel(pydantic.BaseModel):
 
 def _session_with_output(fmt: Any | None) -> RealtimeSessionCreateRequest:
     if fmt is None:
-        return RealtimeSessionCreateRequest(type="realtime", model="gpt-realtime")
+        return RealtimeSessionCreateRequest(type="realtime", model="gpt-realtime-1.5")
     return RealtimeSessionCreateRequest(
         type="realtime",
-        model="gpt-realtime",
+        model="gpt-realtime-1.5",
         # Use dict for output to avoid importing non-exported symbols in tests
         audio=RealtimeAudioConfig(output=cast(Any, {"format": fmt})),
     )
@@ -49,7 +49,7 @@ def test_normalize_session_payload_variants() -> None:
     assert Model._normalize_session_payload(transcription_mapping) is None
 
     # Valid realtime mapping should be converted to model
-    realtime_mapping: Mapping[str, object] = {"type": "realtime", "model": "gpt-realtime"}
+    realtime_mapping: Mapping[str, object] = {"type": "realtime", "model": "gpt-realtime-1.5"}
     as_model = Model._normalize_session_payload(realtime_mapping)
     assert isinstance(as_model, RealtimeSessionCreateRequest)
     assert as_model.type == "realtime"

tests/realtime/test_tracing.py

@@ -100,7 +100,11 @@ async def async_websocket(*args, **kwargs):
                 session_created_event = {
                     "type": "session.created",
                     "event_id": "event_123",
-                    "session": {"id": "session_456", "type": "realtime", "model": "gpt-realtime"},
+                    "session": {
+                        "id": "session_456",
+                        "type": "realtime",
+                        "model": "gpt-realtime-1.5",
+                    },
                 }
 
                 with patch.object(model, "_send_raw_message") as mock_send_raw_message:
@@ -141,7 +145,11 @@ async def async_websocket(*args, **kwargs):
                 session_created_event = {
                     "type": "session.created",
                     "event_id": "event_123",
-                    "session": {"id": "session_456", "type": "realtime", "model": "gpt-realtime"},
+                    "session": {
+                        "id": "session_456",
+                        "type": "realtime",
+                        "model": "gpt-realtime-1.5",
+                    },
                 }
 
                 with patch.object(model, "_send_raw_message") as mock_send_raw_message:
@@ -166,7 +174,7 @@ async def test_tracing_config_none_skips_session_update(self, model, mock_websoc
         session_created_event = {
             "type": "session.created",
             "event_id": "event_123",
-            "session": {"id": "session_456", "type": "realtime", "model": "gpt-realtime"},
+            "session": {"id": "session_456", "type": "realtime", "model": "gpt-realtime-1.5"},
         }
 
         with patch.object(model, "send_event") as mock_send_event:
@@ -205,7 +213,11 @@ async def async_websocket(*args, **kwargs):
                 session_created_event = {
                     "type": "session.created",
                     "event_id": "event_123",
-                    "session": {"id": "session_456", "type": "realtime", "model": "gpt-realtime"},
+                    "session": {
+                        "id": "session_456",
+                        "type": "realtime",
+                        "model": "gpt-realtime-1.5",
+                    },
                 }
 
                 with patch.object(model, "_send_raw_message") as mock_send_raw_message: