diff --git a/.sampo/changesets/cumulative-delta-output-tokens.md b/.sampo/changesets/cumulative-delta-output-tokens.md
new file mode 100644
index 000000000..cbc8bc720
--- /dev/null
+++ b/.sampo/changesets/cumulative-delta-output-tokens.md
@@ -0,0 +1,5 @@
+---
+pypi/posthog: patch
+---
+
+Streamed Anthropic generations no longer stack output tokens when a stream emits more than one `message_delta` (server tools can send several). Each delta's usage is cumulative, so the accumulator now keeps the latest total instead of adding every delta together.
diff --git a/posthog/ai/anthropic/_anthropic_stream.py b/posthog/ai/anthropic/_anthropic_stream.py
index ff3fbe0b8..2df160ef8 100644
--- a/posthog/ai/anthropic/_anthropic_stream.py
+++ b/posthog/ai/anthropic/_anthropic_stream.py
@@ -53,7 +53,15 @@ def consume(self, event: Any) -> None:
         event_usage = extract_anthropic_usage_from_event(event)
         merge_usage_stats(self.usage_stats, event_usage)
         if getattr(event, "type", None) == "message_delta":
-            self._apply_cumulative_input_usage(getattr(event, "usage", None))
+            usage = getattr(event, "usage", None)
+            self._apply_cumulative_input_usage(usage)
+            output_tokens = getattr(usage, "output_tokens", None)
+            if isinstance(output_tokens, int) and not isinstance(output_tokens, bool):
+                # output_tokens on a message_delta is cumulative too, so the
+                # incremental merge above must not stack successive deltas
+                # (server tools can emit several in one stream). Keep the
+                # latest total; a field the event omits leaves the prior one.
+                self.usage_stats["output_tokens"] = output_tokens
 
         if getattr(event, "type", None) == "content_block_start":
             block, tool = handle_anthropic_content_block_start(event)
diff --git a/posthog/test/ai/anthropic/test_anthropic.py b/posthog/test/ai/anthropic/test_anthropic.py
index 08c66e7e2..ffc6feeed 100644
--- a/posthog/test/ai/anthropic/test_anthropic.py
+++ b/posthog/test/ai/anthropic/test_anthropic.py
@@ -907,6 +907,87 @@ async def mock_async_create(**kwargs):
     assert_cumulative_delta_usage(mock_client.capture.call_args.kwargs["properties"])
 
 
+def anthropic_stream_with_two_cumulative_deltas():
+    # Server tools can emit more than one message_delta in a single stream,
+    # and each delta's usage is cumulative: later values replace earlier ones
+    # rather than adding to them.
+    return [
+        RawMessageStartEvent(
+            type="message_start",
+            message=Message(
+                id="msg_1",
+                type="message",
+                role="assistant",
+                model="claude-haiku-4-5",
+                content=[],
+                stop_reason=None,
+                stop_sequence=None,
+                usage=Usage(
+                    input_tokens=100,
+                    output_tokens=1,
+                    cache_creation_input_tokens=0,
+                    cache_read_input_tokens=0,
+                ),
+            ),
+        ),
+        RawMessageDeltaEvent(
+            type="message_delta",
+            delta=Delta(stop_reason="pause_turn", stop_sequence=None),
+            usage=MessageDeltaUsage(output_tokens=50),
+        ),
+        RawMessageDeltaEvent(
+            type="message_delta",
+            delta=Delta(stop_reason="end_turn", stop_sequence=None),
+            usage=MessageDeltaUsage(output_tokens=110),
+        ),
+        RawMessageStopEvent(type="message_stop"),
+    ]
+
+
+def test_streaming_multiple_message_deltas_do_not_stack_output_tokens(mock_client):
+    with patch(
+        "anthropic.resources.Messages.create",
+        return_value=iter(anthropic_stream_with_two_cumulative_deltas()),
+    ):
+        client = Anthropic(api_key="test-key", posthog_client=mock_client)
+        response = client.messages.create(
+            model="claude-haiku-4-5",
+            messages=[{"role": "user", "content": "Hello"}],
+            stream=True,
+        )
+        list(response)
+
+    props = mock_client.capture.call_args.kwargs["properties"]
+    assert props["$ai_output_tokens"] == 110
+
+
+@pytest.mark.asyncio
+async def test_streaming_multiple_message_deltas_do_not_stack_output_tokens_async(
+    mock_client,
+):
+    async def mock_async_stream():
+        for event in anthropic_stream_with_two_cumulative_deltas():
+            yield event
+
+    async def mock_async_create(**kwargs):
+        return mock_async_stream()
+
+    with patch(
+        "anthropic.resources.messages.AsyncMessages.create",
+        side_effect=mock_async_create,
+    ):
+        client = AsyncAnthropic(api_key="test-key", posthog_client=mock_client)
+        response = await client.messages.create(
+            model="claude-haiku-4-5",
+            messages=[{"role": "user", "content": "Hello"}],
+            stream=True,
+        )
+        [event async for event in response]
+
+    props = mock_client.capture.call_args.kwargs["properties"]
+    assert props["$ai_output_tokens"] == 110
+
+
 def test_anthropic_bedrock_preserves_cache_creation_ttl_breakdown(
     mock_client, mock_anthropic_response_with_cache_ttl
 ):