diff --git a/pydantic_ai_slim/pydantic_ai/messages.py b/pydantic_ai_slim/pydantic_ai/messages.py
index c8fc93f64b..95c08773a8 100644
--- a/pydantic_ai_slim/pydantic_ai/messages.py
+++ b/pydantic_ai_slim/pydantic_ai/messages.py
@@ -28,7 +28,7 @@
 from ._utils import generate_tool_call_id as _generate_tool_call_id, now_utc as _now_utc
 from ._warnings import PydanticAIDeprecationWarning
 from .exceptions import UnexpectedModelBehavior
-from .usage import RequestUsage
+from .usage import RequestUsage, RunUsage
 
 if TYPE_CHECKING:
     from .models.instrumented import InstrumentationSettings
@@ -2083,7 +2083,7 @@ class ModelResponse:
 
     _: KW_ONLY
 
-    usage: RequestUsage = field(default_factory=RequestUsage)
+    usage: RequestUsage | RunUsage = field(default_factory=RequestUsage)
     """Usage information for the request.
 
     This has a default to make tests easier, and to support loading old messages where usage will be missing.
diff --git a/pydantic_ai_slim/pydantic_ai/models/wrapper.py b/pydantic_ai_slim/pydantic_ai/models/wrapper.py
index 14591d8df6..5e8c6d45b6 100644
--- a/pydantic_ai_slim/pydantic_ai/models/wrapper.py
+++ b/pydantic_ai_slim/pydantic_ai/models/wrapper.py
@@ -4,7 +4,7 @@
 from contextlib import asynccontextmanager
 from dataclasses import dataclass
 from datetime import datetime
-from typing import Any
+from typing import Any, cast
 
 from typing_extensions import Self
 
@@ -46,7 +46,7 @@ def get(self) -> ModelResponse:
         '`StreamedResponse.usage` is no longer a method; access it as a property (drop the parentheses).'
     )
     def usage(self) -> RequestUsage:
-        return self.response.usage  # pragma: no cover
+        return cast(RequestUsage, self.response.usage)  # pragma: no cover
 
     @property
     def model_name(self) -> str:
diff --git a/tests/models/test_openai.py b/tests/models/test_openai.py
index a6282917a7..0a6e5dba88 100644
--- a/tests/models/test_openai.py
+++ b/tests/models/test_openai.py
@@ -5283,7 +5283,7 @@ async def test_stream_with_continuous_usage_stats(allow_model_requests: None):
         # Verify usage is updated at each step via stream_response
         usage_at_each_step: list[RequestUsage] = []
         async for response in result.stream_response(debounce_by=None):
-            usage_at_each_step.append(response.usage)
+            usage_at_each_step.append(cast(RequestUsage, response.usage))
 
         # Each step should have the cumulative usage from that chunk (not accumulated)
         # The stream emits responses for each content chunk plus final
diff --git a/tests/test_messages.py b/tests/test_messages.py
index 0b46d73142..0cf121078b 100644
--- a/tests/test_messages.py
+++ b/tests/test_messages.py
@@ -24,6 +24,7 @@
     NativeToolReturnPart,
     RequestUsage,
     RetryPromptPart,
+    RunUsage,
     TextContent,
     TextPart,
     ThinkingPart,
@@ -509,6 +510,26 @@ def test_pre_usage_refactor_messages_deserializable():
     )
 
 
+def test_model_response_preserves_run_usage_roundtrip():
+    response = ModelResponse(
+        parts=[TextPart(content='Hello!')],
+        usage=RunUsage(requests=5, tool_calls=3, input_tokens=1000, output_tokens=500),
+        model_name='test',
+    )
+
+    serialized = ModelMessagesTypeAdapter.dump_json([response])
+    deserialized = ModelMessagesTypeAdapter.validate_json(serialized)
+
+    msg = deserialized[0]
+    assert isinstance(msg, ModelResponse)
+    usage = msg.usage
+    assert isinstance(usage, RunUsage)
+    assert usage.requests == 5
+    assert usage.tool_calls == 3
+    assert usage.input_tokens == 1000
+    assert usage.output_tokens == 500
+
+
 def test_file_part_has_content():
     filepart = FilePart(content=BinaryContent(data=b'', media_type='application/pdf'))
     assert not filepart.has_content()