fix(anthropic): take cumulative usage from message_delta when streaming

CHANGELOG.md

@@ -27,6 +27,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 * `ChatOllama()` can now be created without a `model`, so `ChatOllama().list_models()` works. Sending a message without a model still raises an error that lists the available models. (#166)
 * `ChatGoogle()` now reports `finish_reason == "content_filter"` for Gemini's `IMAGE_SAFETY`, `IMAGE_PROHIBITED_CONTENT`, `IMAGE_RECITATION` and `LANGUAGE` finish reasons, instead of passing the raw Gemini string through, and reports `"tool_use"` (rather than `"success"`) when a turn ends with a tool request, as `ChatOpenAI()` does. (#473)
 * The error raised for a tool request with invalid JSON arguments from `ChatOpenAI()` and `ChatOpenAICompletions()` (and the providers built on them) now shows the offending arguments, instead of the literal text `{args}`.
+* When streaming, `ChatAnthropic()` (and `ChatBedrockAnthropic()`) now takes the input and cache token counts from the final `message_delta` event, which reports cumulative usage, as the non-streaming response does. With server tools such as web search, those counts grow during the response, so streamed turns under-reported their tokens and cost (e.g. missing the cache writes made while searching). (#466)
 * `content_image_url()` with a `data:` URL no longer silently drops the `detail` argument: the setting is now carried by inline image content and forwarded to OpenAI's Responses and Chat Completions APIs (regular image URLs already forwarded it). Other providers continue to ignore `detail`. (#467)
 * `ChatBedrock()` on a Converse model no longer drops redacted reasoning blocks when streaming. They were discarded as they arrived, so a streamed turn lost them, while a non-streamed one kept them. (#468)
 * `ChatBedrockAnthropic()` now uses native structured outputs for `data_model` on Claude 4.5 and later, as `ChatAnthropic()` does. Its Bedrock model IDs (`anthropic.claude-...`, `us.anthropic.claude-...`) weren't recognized, so it fell back to forcing a tool call, which Claude Opus 5.5 and Sonnet 5.5 reject with a 400.

chatlas/_provider_anthropic.py

@@ -704,6 +704,18 @@ def stream_merge_chunks(self, completion, chunk):
             completion.stop_reason = chunk.delta.stop_reason
             completion.stop_sequence = chunk.delta.stop_sequence
             completion.usage.output_tokens = chunk.usage.output_tokens
+            # Usage on message_delta is cumulative and, with server tools, can
+            # exceed message_start's input and cache counts. Fields are omitted
+            # when not applicable, in which case message_start's value stands.
+            for field in (
+                "input_tokens",
+                "cache_creation_input_tokens",
+                "cache_read_input_tokens",
+                "server_tool_use",
+            ):
+                value = getattr(chunk.usage, field, None)
+                if value is not None:
+                    setattr(completion.usage, field, value)
 
         return completion
 

tests/test_provider_anthropic.py

@@ -9,10 +9,13 @@
     CitationsWebSearchResultLocation,
     DocumentBlock,
     Message,
+    MessageDeltaUsage,
     RawContentBlockDeltaEvent,
     RawContentBlockStartEvent,
     RawContentBlockStopEvent,
+    RawMessageDeltaEvent,
     RawMessageStartEvent,
+    ServerToolUsage,
     TextBlock,
     TextDelta,
     Usage,
@@ -302,6 +305,93 @@ def test_anthropic_web_fetch_citation_uses_document_index_for_source():
     assert citation.source.title == "Second document"
 
 
+def test_anthropic_stream_usage_takes_cumulative_message_delta_counts():
+    # With server tools, message_delta carries the cumulative input and cache
+    # counts, which differ from message_start (values from the
+    # test_anthropic_web_search_streaming recording).
+    provider = AnthropicProvider(model="claude-haiku-4-5", api_key="dummy", kwargs=None)
+    start_usage = Usage(
+        input_tokens=2247,
+        cache_creation_input_tokens=0,
+        cache_read_input_tokens=0,
+        output_tokens=26,
+    )
+    start = RawMessageStartEvent(
+        type="message_start",
+        message=Message(
+            id="msg",
+            type="message",
+            role="assistant",
+            model="claude-haiku-4-5",
+            content=[],
+            stop_reason=None,
+            stop_sequence=None,
+            usage=start_usage,
+        ),
+    )
+    delta = RawMessageDeltaEvent.model_validate(
+        {
+            "type": "message_delta",
+            "delta": {"stop_reason": "end_turn", "stop_sequence": None},
+            "usage": MessageDeltaUsage(
+                input_tokens=2254,
+                cache_creation_input_tokens=17269,
+                cache_read_input_tokens=0,
+                output_tokens=110,
+                server_tool_use=ServerToolUsage(
+                    web_search_requests=1, web_fetch_requests=0
+                ),
+            ).model_dump(),
+        }
+    )
+
+    completion = provider.stream_merge_chunks(None, start)
+    completion = provider.stream_merge_chunks(completion, delta)
+
+    assert completion.usage.input_tokens == 2254
+    assert completion.usage.cache_creation_input_tokens == 17269
+    assert completion.usage.cache_read_input_tokens == 0
+    assert completion.usage.output_tokens == 110
+    assert completion.usage.server_tool_use == ServerToolUsage(
+        web_search_requests=1, web_fetch_requests=0
+    )
+    assert provider.value_tokens(completion) == (2254 + int(17269 * 1.25), 110, 0)
+
+
+def test_anthropic_stream_usage_keeps_message_start_counts_omitted_by_delta():
+    provider = AnthropicProvider(model="claude-haiku-4-5", api_key="dummy", kwargs=None)
+    start = RawMessageStartEvent(
+        type="message_start",
+        message=Message(
+            id="msg",
+            type="message",
+            role="assistant",
+            model="claude-haiku-4-5",
+            content=[],
+            stop_reason=None,
+            stop_sequence=None,
+            usage=Usage(
+                input_tokens=3,
+                cache_creation_input_tokens=37,
+                cache_read_input_tokens=17269,
+                output_tokens=1,
+            ),
+        ),
+    )
+    delta = RawMessageDeltaEvent.model_validate(
+        {
+            "type": "message_delta",
+            "delta": {"stop_reason": "end_turn", "stop_sequence": None},
+            "usage": {"output_tokens": 10},
+        }
+    )
+
+    completion = provider.stream_merge_chunks(None, start)
+    completion = provider.stream_merge_chunks(completion, delta)
+
+    assert provider.value_tokens(completion) == (3 + int(37 * 1.25), 10, 17269)
+
+
 def test_anthropic_web_fetch_citation_stream_uses_document_index_for_source():
     provider = AnthropicProvider(
         model="claude-sonnet-4-5", api_key="dummy", kwargs=None
@@ -1166,9 +1256,7 @@ def test_anthropic_message_params_convert_unsigned_thinking_blocks():
 
     turns = [
         UserTurn("hi"),
-        AssistantTurn(
-            [ContentThinking(thinking="hmm"), ContentText(text="hello")]
-        ),
+        AssistantTurn([ContentThinking(thinking="hmm"), ContentText(text="hello")]),
         UserTurn("again"),
     ]
     messages = provider._as_message_params(turns)