fix(anthropic): preserve message cache control for list content

litellm/litellm_core_utils/prompt_templates/factory.py

@@ -2407,6 +2407,7 @@ def anthropic_messages_pt(
             ) = messages[msg_i]
             if user_message_types_block["role"] == "user":
                 if isinstance(user_message_types_block["content"], list):
+                    user_content_start: Final = len(user_content)
                     for m in user_message_types_block["content"]:
                         if m.get("type", "") == "image_url":
                             m = cast(ChatCompletionImageObject, m)
@@ -2478,6 +2479,16 @@ def anthropic_messages_pt(
                                     _file_content_element,
                                 )
                             )
+                    if (
+                        len(user_content) > user_content_start
+                        and user_message_types_block.get("cache_control") is not None
+                        and not any(
+                            m.get("cache_control") is not None
+                            for m in user_message_types_block["content"]
+                            if isinstance(m, dict)
+                        )
+                    ):
+                        add_cache_control_to_content(user_content[-1], dict(user_message_types_block))
                 elif isinstance(user_message_types_block["content"], str):
                     _anthropic_content_text_element: AnthropicMessagesTextParam = {
                         "type": "text",
@@ -2680,6 +2691,7 @@ def anthropic_messages_pt(
                 ):  # IMPORTANT: ADD THIS FIRST, ELSE ANTHROPIC WILL RAISE AN ERROR
                     assistant_content.extend(thinking_blocks)
                 if _content_is_list and _content_list is not None:
+                    assistant_content_start: Final = len(assistant_content)
                     for m in _content_list:
                         if not isinstance(m, dict):
                             continue
@@ -2710,6 +2722,12 @@ def anthropic_messages_pt(
                         # Pass through as-is since these are Anthropic-native content types
                         elif m.get("type", "") == "server_tool_use" or m.get("type", "").endswith("_tool_result"):
                             assistant_content.append(m)
+                    if (
+                        len(assistant_content) > assistant_content_start
+                        and assistant_content_block.get("cache_control") is not None
+                        and not any(m.get("cache_control") is not None for m in _content_list if isinstance(m, dict))
+                    ):
+                        add_cache_control_to_content(assistant_content[-1], dict(assistant_content_block))
                 elif (
                     "content" in assistant_content_block
                     and isinstance(assistant_content_block["content"], str)
@@ -4417,6 +4435,16 @@ class BedrockConverseMessagesProcessor:
                             )
                             if _cache_point_block is not None:
                                 _parts.append(_cache_point_block)
+                    if (
+                        _parts
+                        and message_block.get("cache_control") is not None
+                        and not any("cachePoint" in part for part in _parts)
+                    ):
+                        _cache_point_block = litellm.AmazonConverseConfig().get_cache_point_block(
+                            message_block, block_type="content_block", model=model
+                        )
+                        if _cache_point_block is not None:
+                            _parts.append(_cache_point_block)
                     user_content.extend(_parts)
                 elif message_block["content"] and isinstance(message_block["content"], str):
                     _part = BedrockContentBlock(text=messages[msg_i]["content"])
@@ -4790,6 +4818,16 @@ def _bedrock_converse_messages_pt(
                         )
                         if _cache_point_block is not None:
                             _parts.append(_cache_point_block)
+                if (
+                    _parts
+                    and message_block.get("cache_control") is not None
+                    and not any("cachePoint" in part for part in _parts)
+                ):
+                    _cache_point_block = litellm.AmazonConverseConfig().get_cache_point_block(
+                        message_block, block_type="content_block", model=model
+                    )
+                    if _cache_point_block is not None:
+                        _parts.append(_cache_point_block)
                 user_content.extend(_parts)
             elif message_block["content"] and isinstance(message_block["content"], str):
                 _part = BedrockContentBlock(text=messages[msg_i]["content"])

litellm/llms/anthropic/chat/transformation.py

@@ -1719,6 +1719,7 @@ class AnthropicConfig(AnthropicModelInfo, BaseConfig):
                         anthropic_system_message_content["cache_control"] = system_message_block["cache_control"]
                     anthropic_system_message_list.append(anthropic_system_message_content)
                 elif isinstance(message["content"], list):
+                    content_start: Final = len(anthropic_system_message_list)
                     for _content in message["content"]:
                         # Skip empty text blocks - Anthropic API raises errors for empty text
                         text_value = _content.get("text")
@@ -1739,6 +1740,16 @@ class AnthropicConfig(AnthropicModelInfo, BaseConfig):
                             anthropic_system_message_content["cache_control"] = _content["cache_control"]
 
                         anthropic_system_message_list.append(anthropic_system_message_content)
+                    if (
+                        len(anthropic_system_message_list) > content_start
+                        and system_message_block.get("cache_control") is not None
+                        and not any(
+                            content.get("cache_control") is not None
+                            for content in message["content"]
+                            if isinstance(content, dict)
+                        )
+                    ):
+                        anthropic_system_message_list[-1]["cache_control"] = system_message_block["cache_control"]
 
         if len(system_prompt_indices) > 0:
             for idx in reversed(system_prompt_indices):

tests/unit/litellm_core_utils/prompt_templates/test_litellm_core_utils_prompt_templates_factory.py

@@ -30,6 +30,56 @@ from litellm.litellm_core_utils.prompt_templates.factory import (
 from litellm.types.llms.openai import ChatCompletionToolMessage
 
 
+@pytest.mark.parametrize("role", ["user", "assistant"])
+@pytest.mark.parametrize("block_cache_control", [False, True])
+def test_anthropic_list_content_preserves_message_cache_control(role, block_cache_control):
+    message_cache_control: Final = {"type": "ephemeral"}
+    block_cache: Final = {"type": "ephemeral", "ttl": "1h"}
+    content: Final = [
+        {"type": "text", "text": "first", **({"cache_control": block_cache} if block_cache_control else {})},
+        {"type": "text", "text": "second"},
+    ]
+    message: Final = {"role": role, "content": content, "cache_control": message_cache_control}
+    messages: Final = [message] if role == "user" else [{"role": "user", "content": "hi"}, message]
+
+    result: Final = anthropic_messages_pt(messages=messages, model="claude-sonnet-4-6", llm_provider="anthropic")
+
+    expected: Final = [
+        {"type": "text", "text": "first", **({"cache_control": block_cache} if block_cache_control else {})},
+        {
+            "type": "text",
+            "text": "second",
+            **({} if block_cache_control else {"cache_control": message_cache_control}),
+        },
+    ]
+    assert result[-1]["content"] == expected
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("block_cache_control", [False, True])
+async def test_bedrock_list_content_preserves_message_cache_control(block_cache_control):
+    content: Final = [
+        {"type": "text", "text": "first", **({"cache_control": {"type": "ephemeral"}} if block_cache_control else {})},
+        {"type": "text", "text": "second"},
+    ]
+    messages: Final = [{"role": "user", "content": content, "cache_control": {"type": "ephemeral"}}]
+    model: Final = "us.anthropic.claude-sonnet-4-5-20250929-v1:0"
+
+    sync_result: Final = _bedrock_converse_messages_pt(messages=messages, model=model, llm_provider="bedrock")
+    async_result: Final = await BedrockConverseMessagesProcessor._bedrock_converse_messages_pt_async(
+        messages=messages, model=model, llm_provider="bedrock"
+    )
+
+    cache_point: Final = {"cachePoint": {"type": "default"}}
+    expected: Final = (
+        [{"text": "first"}, cache_point, {"text": "second"}]
+        if block_cache_control
+        else [{"text": "first"}, {"text": "second"}, cache_point]
+    )
+    assert sync_result[0]["content"] == expected
+    assert async_result[0]["content"] == expected
+
+
 def _get_gemini_function_response_inline_data_parts(result):
     assert isinstance(result, list), "expected Gemini parts list"
     assert len(result) == 1, "multimodal function responses should stay in one part"

tests/unit/llms/anthropic/chat/test_anthropic_chat_transformation.py

@@ -2383,6 +2383,32 @@ def test_translate_system_message_preserves_cache_control():
     assert result[0]["cache_control"] == {"type": "ephemeral"}
 
 
+@pytest.mark.parametrize("block_cache_control", [False, True])
+def test_translate_system_list_content_preserves_message_cache_control(block_cache_control):
+    message_cache_control: Final = {"type": "ephemeral"}
+    block_cache: Final = {"type": "ephemeral", "ttl": "1h"}
+    content: Final = [
+        {"type": "text", "text": "first", **({"cache_control": block_cache} if block_cache_control else {})},
+        {"type": "text", "text": "second"},
+    ]
+    messages: Final = [
+        {"role": "system", "content": content, "cache_control": message_cache_control},
+        {"role": "user", "content": "hi"},
+    ]
+
+    result: Final = AnthropicConfig().translate_system_message(messages)
+
+    expected: Final = [
+        {"type": "text", "text": "first", **({"cache_control": block_cache} if block_cache_control else {})},
+        {
+            "type": "text",
+            "text": "second",
+            **({} if block_cache_control else {"cache_control": message_cache_control}),
+        },
+    ]
+    assert result == expected
+
+
 # ============ Dynamic max_tokens Tests ============