fix(anthropic): correct input_token count for streaming (#32591)

* Create usage metadata on
[`message_delta`](https://docs.anthropic.com/en/docs/build-with-claude/streaming#event-types)
instead of at the beginning. Consequently, token counts are not included
during streaming but instead at the end. This allows for accurate
reporting of server-side tool usage (important for billing)
* Add some clarifying comments
* Fix some outstanding Pylance warnings
* Remove unnecessary `text` popping in thinking blocks
* Also now correctly reports `input_cache_read`/`input_cache_creation`
as a result

libs/partners/anthropic/langchain_anthropic/chat_models.py

@@ -2192,47 +2192,65 @@ def _make_message_chunk_from_anthropic_event(
     coerce_content_to_string: bool,
     block_start_event: Optional[anthropic.types.RawMessageStreamEvent] = None,
 ) -> tuple[Optional[AIMessageChunk], Optional[anthropic.types.RawMessageStreamEvent]]:
-    """Convert Anthropic event to AIMessageChunk.
+    """Convert Anthropic streaming event to `AIMessageChunk`.
+
+    Args:
+        event: Raw streaming event from Anthropic SDK
+        stream_usage: Whether to include usage metadata in the output chunks.
+        coerce_content_to_string: Whether to convert structured content to plain
+            text strings. When True, only text content is preserved; when False,
+            structured content like tool calls and citations are maintained.
+        block_start_event: Previous content block start event, used for tracking
+            tool use blocks and maintaining context across related events.
+
+    Returns:
+        Tuple containing:
+        - AIMessageChunk: Converted message chunk with appropriate content and
+          metadata, or None if the event doesn't produce a chunk
+        - RawMessageStreamEvent: Updated `block_start_event` for tracking content
+          blocks across sequential events, or None if not applicable
+
+    Note:
+        Not all Anthropic events result in message chunks. Events like internal
+        state changes return None for the message chunk while potentially
+        updating the `block_start_event` for context tracking.
 
-    Note that not all events will result in a message chunk. In these cases
-    we return ``None``.
     """
     message_chunk: Optional[AIMessageChunk] = None
-    # See https://github.com/anthropics/anthropic-sdk-python/blob/main/src/anthropic/lib/streaming/_messages.py  # noqa: E501
+    # Reference: Anthropic SDK streaming implementation
+    # https://github.com/anthropics/anthropic-sdk-python/blob/main/src/anthropic/lib/streaming/_messages.py  # noqa: E501
+
     if event.type == "message_start" and stream_usage:
-        usage_metadata = _create_usage_metadata(event.message.usage)
+        # Capture model name, but don't include usage_metadata yet
-        # We pick up a cumulative count of output_tokens at the end of the stream,
+        # as it will be properly reported in message_delta with complete info
-        # so here we zero out to avoid double counting.
-        usage_metadata["total_tokens"] = (
-            usage_metadata["total_tokens"] - usage_metadata["output_tokens"]
-        )
-        usage_metadata["output_tokens"] = 0
         if hasattr(event.message, "model"):
             response_metadata = {"model_name": event.message.model}
         else:
             response_metadata = {}
+
         message_chunk = AIMessageChunk(
             content="" if coerce_content_to_string else [],
-            usage_metadata=usage_metadata,
             response_metadata=response_metadata,
         )
+
     elif (
         event.type == "content_block_start"
         and event.content_block is not None
         and event.content_block.type
         in (
-            "tool_use",
+            "tool_use",  # Standard tool usage
-            "code_execution_tool_result",
+            "code_execution_tool_result",  # Built-in code execution results
             "document",
             "redacted_thinking",
             "mcp_tool_use",
             "mcp_tool_result",
-            "server_tool_use",
+            "server_tool_use",  # Server-side tool usage
-            "web_search_tool_result",
+            "web_search_tool_result",  # Built-in web search results
         )
     ):
         if coerce_content_to_string:
             warnings.warn("Received unexpected tool content block.", stacklevel=2)
+
         content_block = event.content_block.model_dump()
         content_block["index"] = event.index
         if event.content_block.type == "tool_use":
@@ -2250,35 +2268,47 @@ def _make_message_chunk_from_anthropic_event(
             tool_call_chunks=tool_call_chunks,
         )
         block_start_event = event
+
+    # Process incremental content updates
     elif event.type == "content_block_delta":
+        # Text and citation deltas (incremental text content)
         if event.delta.type in ("text_delta", "citations_delta"):
             if coerce_content_to_string and hasattr(event.delta, "text"):
-                text = event.delta.text
+                text = getattr(event.delta, "text", "")
                 message_chunk = AIMessageChunk(content=text)
             else:
                 content_block = event.delta.model_dump()
                 content_block["index"] = event.index
+
+                # All citation deltas are part of a text block
                 content_block["type"] = "text"
                 if "citation" in content_block:
+                    # Assign citations to a list if present
                     content_block["citations"] = [content_block.pop("citation")]
                 message_chunk = AIMessageChunk(content=[content_block])
+
+        # Reasoning
         elif (
             event.delta.type == "thinking_delta"
             or event.delta.type == "signature_delta"
         ):
             content_block = event.delta.model_dump()
-            if "text" in content_block and content_block["text"] is None:
-                content_block.pop("text")
             content_block["index"] = event.index
             content_block["type"] = "thinking"
             message_chunk = AIMessageChunk(content=[content_block])
+
+        # Tool input JSON (streaming tool arguments)
         elif event.delta.type == "input_json_delta":
             content_block = event.delta.model_dump()
             content_block["index"] = event.index
+            start_event_block = (
+                getattr(block_start_event, "content_block", None)
+                if block_start_event
+                else None
+            )
             if (
-                (block_start_event is not None)
+                start_event_block is not None
-                and hasattr(block_start_event, "content_block")
+                and getattr(start_event_block, "type", None) == "tool_use"
-                and (block_start_event.content_block.type == "tool_use")
             ):
                 tool_call_chunk = create_tool_call_chunk(
                     index=event.index,
@@ -2293,12 +2323,10 @@ def _make_message_chunk_from_anthropic_event(
                 content=[content_block],
                 tool_call_chunks=tool_call_chunks,
             )
+
+    # Process final usage metadata and completion info
     elif event.type == "message_delta" and stream_usage:
-        usage_metadata = UsageMetadata(
+        usage_metadata = _create_usage_metadata(event.usage)
-            input_tokens=0,
-            output_tokens=event.usage.output_tokens,
-            total_tokens=event.usage.output_tokens,
-        )
         message_chunk = AIMessageChunk(
             content="",
             usage_metadata=usage_metadata,
@@ -2307,6 +2335,8 @@ def _make_message_chunk_from_anthropic_event(
                 "stop_sequence": event.delta.stop_sequence,
             },
         )
+    # Unhandled event types (e.g., `content_block_stop`, `ping` events)
+    # https://docs.anthropic.com/en/docs/build-with-claude/streaming#other-events
     else:
         pass
 
@@ -2319,26 +2349,38 @@ class ChatAnthropicMessages(ChatAnthropic):
 
 
 def _create_usage_metadata(anthropic_usage: BaseModel) -> UsageMetadata:
+    """Create LangChain `UsageMetadata` from Anthropic `Usage` data.
+
+    Note: Anthropic's `input_tokens` excludes cached tokens, so we manually add
+    `cache_read` and `cache_creation` tokens to get the true total.
+
+    """
     input_token_details: dict = {
         "cache_read": getattr(anthropic_usage, "cache_read_input_tokens", None),
         "cache_creation": getattr(anthropic_usage, "cache_creation_input_tokens", None),
     }
-    # Add (beta) cache TTL information if available
+
+    # Add cache TTL information if provided (5-minute and 1-hour ephemeral cache)
     cache_creation = getattr(anthropic_usage, "cache_creation", None)
-    cache_creation_keys = ("ephemeral_1h_input_tokens", "ephemeral_5m_input_tokens")
+
+    # Currently just copying over the 5m and 1h keys, but if more are added in the
+    # future we'll need to expand this tuple
+    cache_creation_keys = ("ephemeral_5m_input_tokens", "ephemeral_1h_input_tokens")
     if cache_creation:
         if isinstance(cache_creation, BaseModel):
             cache_creation = cache_creation.model_dump()
         for k in cache_creation_keys:
             input_token_details[k] = cache_creation.get(k)
 
-    # Anthropic input_tokens exclude cached token counts.
+    # Calculate total input tokens: Anthropic's `input_tokens` excludes cached tokens,
+    # so we need to add them back to get the true total input token count
     input_tokens = (
-        (getattr(anthropic_usage, "input_tokens", 0) or 0)
+        (getattr(anthropic_usage, "input_tokens", 0) or 0)  # Base input tokens
-        + (input_token_details["cache_read"] or 0)
+        + (input_token_details["cache_read"] or 0)  # Tokens read from cache
-        + (input_token_details["cache_creation"] or 0)
+        + (input_token_details["cache_creation"] or 0)  # Tokens used to create cache
     )
     output_tokens = getattr(anthropic_usage, "output_tokens", 0) or 0
+
     return UsageMetadata(
         input_tokens=input_tokens,
         output_tokens=output_tokens,

libs/partners/anthropic/pyproject.toml

@@ -7,7 +7,7 @@ authors = []
 license = { text = "MIT" }
 requires-python = ">=3.9"
 dependencies = [
-    "anthropic<1,>=0.60.0",
+    "anthropic<1,>=0.64.0",
     "langchain-core<1.0.0,>=0.3.72",
     "pydantic<3.0.0,>=2.7.4",
 ]

libs/partners/anthropic/tests/integration_tests/test_llms.py

@@ -9,6 +9,8 @@ from langchain_core.outputs import LLMResult
 from langchain_anthropic import Anthropic
 from tests.unit_tests._utils import FakeCallbackHandler
 
+MODEL = "claude-3-7-sonnet-latest"
+
 
 @pytest.mark.requires("anthropic")
 def test_anthropic_model_name_param() -> None:
@@ -24,14 +26,14 @@ def test_anthropic_model_param() -> None:
 
 def test_anthropic_call() -> None:
     """Test valid call to anthropic."""
-    llm = Anthropic(model="claude-3-7-sonnet-20250219")  # type: ignore[call-arg]
+    llm = Anthropic(model=MODEL)  # type: ignore[call-arg]
     output = llm.invoke("Say foo:")
     assert isinstance(output, str)
 
 
 def test_anthropic_streaming() -> None:
     """Test streaming tokens from anthropic."""
-    llm = Anthropic(model="claude-3-7-sonnet-20250219")  # type: ignore[call-arg]
+    llm = Anthropic(model=MODEL)  # type: ignore[call-arg]
     generator = llm.stream("I'm Pickle Rick")
 
     assert isinstance(generator, Generator)
@@ -45,7 +47,7 @@ def test_anthropic_streaming_callback() -> None:
     callback_handler = FakeCallbackHandler()
     callback_manager = CallbackManager([callback_handler])
     llm = Anthropic(
-        model="claude-3-7-sonnet-20250219",  # type: ignore[call-arg]
+        model=MODEL,  # type: ignore[call-arg]
         streaming=True,
         callback_manager=callback_manager,
         verbose=True,
@@ -56,7 +58,7 @@ def test_anthropic_streaming_callback() -> None:
 
 async def test_anthropic_async_generate() -> None:
     """Test async generate."""
-    llm = Anthropic(model="claude-3-7-sonnet-20250219")  # type: ignore[call-arg]
+    llm = Anthropic(model=MODEL)  # type: ignore[call-arg]
     output = await llm.agenerate(["How many toes do dogs have?"])
     assert isinstance(output, LLMResult)
 
@@ -66,7 +68,7 @@ async def test_anthropic_async_streaming_callback() -> None:
     callback_handler = FakeCallbackHandler()
     callback_manager = CallbackManager([callback_handler])
     llm = Anthropic(
-        model="claude-3-7-sonnet-20250219",  # type: ignore[call-arg]
+        model=MODEL,  # type: ignore[call-arg]
         streaming=True,
         callback_manager=callback_manager,
         verbose=True,

libs/partners/anthropic/tests/integration_tests/test_standard.py

@@ -11,6 +11,8 @@ from langchain_anthropic import ChatAnthropic
 
 REPO_ROOT_DIR = Path(__file__).parents[5]
 
+MODEL = "claude-3-5-haiku-latest"
+
 
 class TestAnthropicStandard(ChatModelIntegrationTests):
     @property
@@ -19,7 +21,7 @@ class TestAnthropicStandard(ChatModelIntegrationTests):
 
     @property
     def chat_model_params(self) -> dict:
-        return {"model": "claude-3-5-sonnet-latest"}
+        return {"model": MODEL}
 
     @property
     def supports_image_inputs(self) -> bool:
@@ -67,8 +69,7 @@ class TestAnthropicStandard(ChatModelIntegrationTests):
 
     def invoke_with_cache_creation_input(self, *, stream: bool = False) -> AIMessage:
         llm = ChatAnthropic(
-            model="claude-3-5-sonnet-20240620",  # type: ignore[call-arg]
+            model=MODEL,  # type: ignore[call-arg]
-            extra_headers={"anthropic-beta": "prompt-caching-2024-07-31"},  # type: ignore[call-arg]
         )
         with open(REPO_ROOT_DIR / "README.md") as f:
             readme = f.read()
@@ -96,8 +97,7 @@ class TestAnthropicStandard(ChatModelIntegrationTests):
 
     def invoke_with_cache_read_input(self, *, stream: bool = False) -> AIMessage:
         llm = ChatAnthropic(
-            model="claude-3-5-sonnet-20240620",  # type: ignore[call-arg]
+            model=MODEL,  # type: ignore[call-arg]
-            extra_headers={"anthropic-beta": "prompt-caching-2024-07-31"},  # type: ignore[call-arg]
         )
         with open(REPO_ROOT_DIR / "README.md") as f:
             readme = f.read()

libs/partners/anthropic/tests/unit_tests/test_chat_models.py

@@ -1213,3 +1213,76 @@ def test_cache_control_kwarg() -> None:
             ],
         },
     ]
+
+
+def test_streaming_cache_token_reporting() -> None:
+    """Test that cache tokens are properly reported in streaming events."""
+    from unittest.mock import MagicMock
+
+    from anthropic.types import MessageDeltaUsage
+
+    from langchain_anthropic.chat_models import _make_message_chunk_from_anthropic_event
+
+    # Create a mock message_start event
+    mock_message = MagicMock()
+    mock_message.model = "claude-3-sonnet-20240229"
+    mock_message.usage.input_tokens = 100
+    mock_message.usage.output_tokens = 0
+    mock_message.usage.cache_read_input_tokens = 25
+    mock_message.usage.cache_creation_input_tokens = 10
+
+    message_start_event = MagicMock()
+    message_start_event.type = "message_start"
+    message_start_event.message = mock_message
+
+    # Create a mock message_delta event with complete usage info
+    mock_delta_usage = MessageDeltaUsage(
+        output_tokens=50,
+        input_tokens=100,
+        cache_read_input_tokens=25,
+        cache_creation_input_tokens=10,
+    )
+
+    mock_delta = MagicMock()
+    mock_delta.stop_reason = "end_turn"
+    mock_delta.stop_sequence = None
+
+    message_delta_event = MagicMock()
+    message_delta_event.type = "message_delta"
+    message_delta_event.usage = mock_delta_usage
+    message_delta_event.delta = mock_delta
+
+    # Test message_start event
+    start_chunk, _ = _make_message_chunk_from_anthropic_event(
+        message_start_event,
+        stream_usage=True,
+        coerce_content_to_string=True,
+        block_start_event=None,
+    )
+
+    # Test message_delta event - should contain complete usage metadata (w/ cache)
+    delta_chunk, _ = _make_message_chunk_from_anthropic_event(
+        message_delta_event,
+        stream_usage=True,
+        coerce_content_to_string=True,
+        block_start_event=None,
+    )
+
+    # Verify message_delta has complete usage_metadata including cache tokens
+    assert start_chunk is not None, "message_start should produce a chunk"
+    assert getattr(start_chunk, "usage_metadata", None) is None, (
+        "message_start should not have usage_metadata"
+    )
+    assert delta_chunk is not None, "message_delta should produce a chunk"
+    assert delta_chunk.usage_metadata is not None, (
+        "message_delta should have usage_metadata"
+    )
+    assert "input_token_details" in delta_chunk.usage_metadata
+    input_details = delta_chunk.usage_metadata["input_token_details"]
+    assert input_details.get("cache_read") == 25
+    assert input_details.get("cache_creation") == 10
+
+    # Verify totals are correct: 100 base + 25 cache_read + 10 cache_creation = 135
+    assert delta_chunk.usage_metadata["input_tokens"] == 135
+    assert delta_chunk.usage_metadata["output_tokens"] == 50
+    assert delta_chunk.usage_metadata["total_tokens"] == 185

libs/partners/anthropic/uv.lock

@@ -1,5 +1,5 @@
 version = 1
-revision = 2
+revision = 3
 requires-python = ">=3.9"
 resolution-markers = [
     "python_full_version >= '3.13' and platform_python_implementation == 'PyPy'",
@@ -469,7 +469,7 @@ typing = [
 
 [package.metadata]
 requires-dist = [
-    { name = "anthropic", specifier = ">=0.60.0,<1" },
+    { name = "anthropic", specifier = ">=0.64.0,<1" },
     { name = "langchain-core", editable = "../../core" },
     { name = "pydantic", specifier = ">=2.7.4,<3.0.0" },
 ]