fix(bedrock): avoid double-counting cache tokens in Anthropic Messages streaming usage

Made-with: Cursor

litellm/llms/bedrock/messages/invoke_transformations/anthropic_claude3_transformation.py

@@ -622,9 +622,9 @@ class AmazonAnthropicClaudeMessagesConfig(
         merges usage from message_start and message_delta but ignores
         message_stop. This method buffers message_delta and, when
         message_stop arrives with cache usage, merges those fields into the
-        message_delta usage and also updates the input_tokens on
-        message_delta to include the full count (uncached + cache_creation +
-        cache_read).
+        message_delta usage. input_tokens is kept as the uncached-only
+        count; downstream calculate_usage adds cache tokens to
+        prompt_tokens.
         """
         _CACHE_FIELDS = ("cache_creation_input_tokens", "cache_read_input_tokens")
         pending_delta = None
@@ -653,12 +653,7 @@ class AmazonAnthropicClaudeMessagesConfig(
 
                 raw_input = stop_usage.get("input_tokens")
                 if raw_input is not None:
-                    uncached = raw_input if isinstance(raw_input, int) else 0
-                    raw_cc = delta_usage.get("cache_creation_input_tokens", 0)
-                    cache_creation = raw_cc if isinstance(raw_cc, int) else 0
-                    raw_cr = delta_usage.get("cache_read_input_tokens", 0)
-                    cache_read = raw_cr if isinstance(raw_cr, int) else 0
-                    delta_usage["input_tokens"] = uncached + cache_creation + cache_read
+                    delta_usage["input_tokens"] = raw_input if isinstance(raw_input, int) else 0
 
                 if delta_usage:
                     pending_delta["usage"] = delta_usage  # type: ignore[arg-type]

tests/test_litellm/llms/bedrock/messages/invoke_transformations/test_anthropic_claude3_transformation.py

@@ -3,6 +3,7 @@ import json
 import os
 import sys
 from datetime import datetime
+from unittest.mock import Mock
 
 import pytest
 
@@ -130,7 +131,8 @@ async def test_bedrock_sse_wrapper_keeps_usage_in_message_start_and_message_delt
     assert "usage" in delta_json
     assert delta_json["usage"]["cache_creation_input_tokens"] == 1562
     assert delta_json["usage"]["cache_read_input_tokens"] == 32392
-    assert delta_json["usage"]["input_tokens"] == 3 + 1562 + 32392
+    assert delta_json["usage"]["input_tokens"] == 3
+    assert delta_json["usage"]["output_tokens"] == 8
 
 
 def test_chunk_parser_usage_transformation():
@@ -599,3 +601,111 @@ def test_bedrock_messages_strips_output_config_with_output_format():
 
     assert "output_config" not in result
     assert "output_format" not in result
+
+
+@pytest.mark.asyncio
+async def test_promote_message_stop_usage_preserves_message_delta_output_tokens():
+    """
+    Bedrock unified /messages streaming can send full usage on message_delta and a
+    conflicting smaller usage on message_stop (e.g. output_tokens 9 vs 12).
+    _promote_message_stop_usage must not replace message_delta output_tokens.
+    """
+    cfg = AmazonAnthropicClaudeMessagesConfig()
+
+    async def _stream():  # type: ignore[return-type]
+        yield {
+            "type": "message_delta",
+            "delta": {"stop_reason": "end_turn", "stop_sequence": None},
+            "usage": {
+                "input_tokens": 3,
+                "cache_creation_input_tokens": 10553,
+                "cache_read_input_tokens": 25490,
+                "output_tokens": 12,
+            },
+        }
+        yield {
+            "type": "message_stop",
+            "usage": {"input_tokens": 3, "output_tokens": 9},
+        }
+
+    merged: list[dict] = []
+    async for chunk in cfg._promote_message_stop_usage(_stream()):
+        if isinstance(chunk, dict):
+            merged.append(chunk)
+
+    assert len(merged) >= 1
+    delta_out = merged[0]
+    assert delta_out["type"] == "message_delta"
+    assert delta_out["usage"]["output_tokens"] == 12
+    assert delta_out["usage"]["cache_creation_input_tokens"] == 10553
+    assert delta_out["usage"]["cache_read_input_tokens"] == 25490
+    assert delta_out["usage"]["input_tokens"] == 3
+
+
+@pytest.mark.asyncio
+async def test_unified_bedrock_messages_sse_usage_and_cost_claude_sonnet_46():
+    """
+    End-to-end for Bedrock Invoke Anthropic Messages (unified) streaming path:
+    dict chunks -> _promote_message_stop_usage -> bedrock_sse_wrapper SSE bytes ->
+    same logging reconstruction as Anthropic /messages. Ensures token counts and
+    completion_cost match model_prices for us.anthropic.claude-sonnet-4-6.
+    """
+    from litellm import completion_cost
+    from litellm.proxy.pass_through_endpoints.llm_provider_handlers.anthropic_passthrough_logging_handler import (
+        AnthropicPassthroughLoggingHandler,
+    )
+
+    cfg = AmazonAnthropicClaudeMessagesConfig()
+
+    async def _stream():  # type: ignore[return-type]
+        yield {
+            "type": "message_delta",
+            "delta": {"stop_reason": "end_turn", "stop_sequence": None},
+            "usage": {
+                "input_tokens": 3,
+                "cache_creation_input_tokens": 10553,
+                "cache_read_input_tokens": 25490,
+                "output_tokens": 12,
+            },
+        }
+        yield {
+            "type": "message_stop",
+            "usage": {"input_tokens": 3, "output_tokens": 9},
+        }
+
+    logging_obj = LiteLLMLoggingObj(
+        model="bedrock/us.anthropic.claude-sonnet-4-6",
+        messages=[{"role": "user", "content": "Hello"}],
+        stream=True,
+        call_type="chat",
+        start_time=datetime.now(),
+        litellm_call_id="test_unified_bedrock_messages_sse_cost",
+        function_id="test_unified_bedrock_messages_sse_cost",
+    )
+
+    collected: list[bytes] = []
+    async for sse in cfg.bedrock_sse_wrapper(
+        completion_stream=_stream(),
+        litellm_logging_obj=logging_obj,
+        request_body={"model": "us.anthropic.claude-sonnet-4-6"},
+    ):
+        collected.append(sse)
+
+    built = AnthropicPassthroughLoggingHandler._build_complete_streaming_response(
+        all_chunks=collected,
+        model="us.anthropic.claude-sonnet-4-6",
+        litellm_logging_obj=Mock(),
+    )
+    assert built.usage is not None
+    assert built.usage.completion_tokens == 12
+    assert built.usage.prompt_tokens == 36046
+    assert built.usage.total_tokens == 36058
+    assert built.usage.cache_creation_input_tokens == 10553
+    assert built.usage.cache_read_input_tokens == 25490
+
+    cost = completion_cost(
+        completion_response=built,
+        model="bedrock/us.anthropic.claude-sonnet-4-6",
+        custom_llm_provider="bedrock",
+    )
+    assert cost == pytest.approx(0.052150725, rel=0, abs=1e-9)