diff --git a/packages/opentelemetry-instrumentation-anthropic/opentelemetry/instrumentation/anthropic/streaming.py b/packages/opentelemetry-instrumentation-anthropic/opentelemetry/instrumentation/anthropic/streaming.py
index bd2d6dc15d..06a6b37fb1 100644
--- a/packages/opentelemetry-instrumentation-anthropic/opentelemetry/instrumentation/anthropic/streaming.py
+++ b/packages/opentelemetry-instrumentation-anthropic/opentelemetry/instrumentation/anthropic/streaming.py
@@ -62,13 +62,12 @@ def _process_response_item(item, complete_response):
             event["finish_reason"] = item.delta.stop_reason
         if item.usage:
             if "usage" in complete_response:
-                item_output_tokens = dict(item.usage).get("output_tokens", 0)
-                existing_output_tokens = complete_response["usage"].get(
-                    "output_tokens", 0
-                )
-                complete_response["usage"]["output_tokens"] = (
-                    item_output_tokens + existing_output_tokens
-                )
+                # message_delta carries the running total for the whole message,
+                # not an increment, so it replaces the partial count reported by
+                # message_start rather than adding to it.
+                item_output_tokens = dict(item.usage).get("output_tokens")
+                if item_output_tokens is not None:
+                    complete_response["usage"]["output_tokens"] = item_output_tokens
             else:
                 complete_response["usage"] = dict(item.usage)
 
diff --git a/packages/opentelemetry-instrumentation-anthropic/tests/test_messages.py b/packages/opentelemetry-instrumentation-anthropic/tests/test_messages.py
index e33a24ac6b..a81e95f783 100644
--- a/packages/opentelemetry-instrumentation-anthropic/tests/test_messages.py
+++ b/packages/opentelemetry-instrumentation-anthropic/tests/test_messages.py
@@ -680,6 +680,7 @@ def test_anthropic_message_streaming_legacy(
     assert text_parts[0]["content"] == response_content
     assert output_messages[-1]["role"] == "assistant"
     assert anthropic_span.attributes[GenAIAttributes.GEN_AI_USAGE_INPUT_TOKENS] == 17
+    assert anthropic_span.attributes[GenAIAttributes.GEN_AI_USAGE_OUTPUT_TOKENS] == 171
     assert (
         anthropic_span.attributes[GenAIAttributes.GEN_AI_USAGE_OUTPUT_TOKENS]
         + anthropic_span.attributes[GenAIAttributes.GEN_AI_USAGE_INPUT_TOKENS]
diff --git a/packages/opentelemetry-instrumentation-anthropic/tests/test_prompt_caching.py b/packages/opentelemetry-instrumentation-anthropic/tests/test_prompt_caching.py
index 82a0254383..4e94d113af 100644
--- a/packages/opentelemetry-instrumentation-anthropic/tests/test_prompt_caching.py
+++ b/packages/opentelemetry-instrumentation-anthropic/tests/test_prompt_caching.py
@@ -759,7 +759,7 @@ def test_anthropic_prompt_caching_stream_legacy(
     assert json.loads(cache_creation_span.attributes[GenAIAttributes.GEN_AI_OUTPUT_MESSAGES])[-1]["role"] == "assistant"
     assert json.loads(cache_read_span.attributes[GenAIAttributes.GEN_AI_OUTPUT_MESSAGES])[-1]["role"] == "assistant"
 
-    _verify_caching_attributes(cache_creation_span, cache_read_span, 1169, 202, 222, 1165)
+    _verify_caching_attributes(cache_creation_span, cache_read_span, 1169, 201, 221, 1165)
 
     # verify metrics
     metrics_data = reader.get_metrics_data()
@@ -820,7 +820,7 @@ def test_anthropic_prompt_caching_stream_with_events_with_content(
     cache_creation_span = spans[0]
     cache_read_span = spans[1]
 
-    _verify_caching_attributes(cache_creation_span, cache_read_span, 1169, 202, 222, 1165)
+    _verify_caching_attributes(cache_creation_span, cache_read_span, 1169, 201, 221, 1165)
 
     # verify metrics
     metrics_data = reader.get_metrics_data()
@@ -986,7 +986,7 @@ def test_anthropic_prompt_caching_stream_with_events_with_no_content(
     cache_creation_span = spans[0]
     cache_read_span = spans[1]
 
-    _verify_caching_attributes(cache_creation_span, cache_read_span, 1169, 202, 222, 1165)
+    _verify_caching_attributes(cache_creation_span, cache_read_span, 1169, 201, 221, 1165)
 
     # verify metrics
     metrics_data = reader.get_metrics_data()
@@ -1089,7 +1089,7 @@ async def test_anthropic_prompt_caching_async_stream_legacy(
     assert cr_input[0]["role"] == "user"
     assert text == cr_input[0]["parts"][0]["content"]
 
-    _verify_caching_attributes(cache_creation_span, cache_read_span, 1171, 290, 257, 1167)
+    _verify_caching_attributes(cache_creation_span, cache_read_span, 1171, 289, 256, 1167)
 
     # verify metrics
     metrics_data = reader.get_metrics_data()
@@ -1151,7 +1151,7 @@ async def test_anthropic_prompt_caching_async_stream_with_events_with_content(
     cache_creation_span = spans[0]
     cache_read_span = spans[1]
 
-    _verify_caching_attributes(cache_creation_span, cache_read_span, 1171, 290, 257, 1167)
+    _verify_caching_attributes(cache_creation_span, cache_read_span, 1171, 289, 256, 1167)
 
     # verify metrics
     metrics_data = reader.get_metrics_data()
@@ -1328,7 +1328,7 @@ async def test_anthropic_prompt_caching_async_stream_with_events_with_no_content
     cache_creation_span = spans[0]
     cache_read_span = spans[1]
 
-    _verify_caching_attributes(cache_creation_span, cache_read_span, 1171, 290, 257, 1167)
+    _verify_caching_attributes(cache_creation_span, cache_read_span, 1171, 289, 256, 1167)
 
     # verify metrics
     metrics_data = reader.get_metrics_data()
@@ -1372,3 +1372,4 @@ def assert_message_in_logs(log: ReadableLogRecord, event_name: str, expected_con
     else:
         assert log.log_record.body
         assert dict(log.log_record.body) == expected_content
+