Merge pull request #41336 from BerriAI/litellm_fix_anthropic_stream_absent_usage

fix(anthropic): tolerate message_delta events without usage when streaming

litellm/llms/anthropic/chat/handler.py

@@ -1167,7 +1167,9 @@ class ModelResponseIterator:
         # (matches OpenAI behavior and non-streaming Anthropic implementation)
         if self.converted_response_format_tool:
             finish_reason = "stop"
-        usage: Final = self._handle_usage(anthropic_usage_chunk=message_delta["usage"])
+        usage: Final = (
+            self._handle_usage(anthropic_usage_chunk=message_delta["usage"]) if "usage" in message_delta else None
+        )
         container: Final = message_delta["delta"].get("container")
         return finish_reason, usage, container
 

litellm/llms/anthropic/experimental_pass_through/adapters/streaming_iterator.py

@@ -376,10 +376,9 @@ class AnthropicStreamWrapper(AdapterCompletionStreamWrapper):
         usage_dict: UsageDelta = LiteLLMAnthropicMessagesAdapter._translate_openai_usage_to_anthropic_usage_delta(
             chunk.usage
         )
-        merged_chunk["usage"] = usage_dict
         if self.applied_edits and "context_management" not in merged_chunk:
             merged_chunk["context_management"] = ContextManagementResponse(applied_edits=list(self.applied_edits))
-        return self._augment_message_delta_usage(merged_chunk)
+        return self._augment_message_delta_usage({**merged_chunk, "usage": usage_dict})
 
     def _handle_choiceless_chunk(self, chunk: "ModelResponseStream") -> bool:
         """Consume an OpenAI-compatible chunk that carries no ``choices``.
@@ -448,8 +447,7 @@ class AnthropicStreamWrapper(AdapterCompletionStreamWrapper):
             }
             iterations.append(message_iteration)
         augmented_usage["iterations"] = iterations
-        augmented["usage"] = augmented_usage
-        return augmented
+        return {**augmented, "usage": augmented_usage}
 
     def _next_compaction_event(self) -> dict[str, object] | None:
         """Return the next compaction content-block SSE event, or ``None``.

litellm/types/llms/anthropic.py

@@ -586,7 +586,7 @@ class MessageBlockDelta(TypedDict):
 
     type: Literal["message_delta"]
     delta: MessageDelta
-    usage: UsageDelta
+    usage: NotRequired[ReadOnly[UsageDelta]]
     context_management: NotRequired[ContextManagementResponse]
 
 

tests/test_litellm/llms/anthropic/chat/test_anthropic_chat_handler.py

@@ -1,6 +1,7 @@
 import json
 import threading
 from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
+from typing import Final
 from unittest.mock import AsyncMock, MagicMock, patch
 
 import httpx
@@ -579,6 +580,20 @@ def test_text_only_streaming_has_index_zero():
             ), f"Expected index=0, got {parsed.choices[0].index}"
 
 
+def test_message_delta_without_usage_returns_chunk_with_no_usage():
+    iterator: Final = ModelResponseIterator(None, sync_stream=True)
+
+    model_response: Final = iterator.chunk_parser(
+        {
+            "type": "message_delta",
+            "delta": {"stop_reason": "end_turn", "stop_sequence": None},
+        }
+    )
+
+    assert model_response.choices[0].finish_reason == "stop"
+    assert model_response.usage is None
+
+
 def test_streaming_thinking_deltas_count_reasoning_tokens_in_usage():
     """Anthropic streaming usage should account for emitted thinking deltas."""
     chunks = [