Preserve cache_control for Qwen on OpenRouter + Anthropic adapter

litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py

@@ -717,7 +717,11 @@ def is_anthropic_claude_model(model: str) -> bool:
         - vertex_ai/*claude* models
         """
         model_lower = model.lower()
-        return "anthropic" in model_lower or "claude" in model_lower
+        return (
+            "anthropic" in model_lower
+            or "claude" in model_lower
+            or "qwen" in model_lower
+        )
 
     @staticmethod
     def translate_thinking_for_model(
@@ -951,9 +955,20 @@ def _add_system_message_to_messages(
             model_name = anthropic_message_request.get("model", "")
             for block in system_content:
                 if isinstance(block, dict) and block.get("type") == "text":
+                    text = block.get("text", "")
+                    # Drop the `x-anthropic-billing-header:` system block.
+                    # Some clients (e.g. Claude Code) inject this block ahead
+                    # of the cache_control marker with a per-request hash; left
+                    # in place it invalidates the upstream prefix-cache key on
+                    # every turn, so cache_read_input_tokens stays 0. The
+                    # sibling messages/transformation.py path already filters it
+                    # via _filter_billing_headers_from_system; this adapter path
+                    # missed it.
+                    if text.startswith("x-anthropic-billing-header:"):
+                        continue
                     text_block: Dict[str, Any] = {
                         "type": "text",
-                        "text": block.get("text", ""),
+                        "text": text,
                     }
                     self._add_cache_control_if_applicable(block, text_block, model_name)
                     openai_system_content.append(text_block)

litellm/llms/openrouter/chat/transformation.py

@@ -30,6 +30,13 @@ class CacheControlSupportedModels(str, Enum):
     MINIMAX = "minimax"
     GLM = "glm"
     ZAI = "z-ai"
+    # OpenRouter's Alibaba (Qwen) upstreams require explicit cache_control
+    # content blocks. See https://openrouter.ai/docs/features/prompt-caching.
+    # Without this the handler strips cache_control before the upstream call,
+    # so every turn pays full input rate (cache_read discount never applies).
+    # DeepSeek is intentionally omitted: its caching is automatic/prefix-based
+    # and ignores cache_control, so listing it here would be inert.
+    QWEN = "qwen"
 
 
 class OpenrouterConfig(OpenAIGPTConfig):

tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_anthropic_experimental_pass_through_adapters_transformation.py

@@ -2472,3 +2472,58 @@ def test_translate_anthropic_tool_choice_none():
 
     result = adapter.translate_anthropic_tool_choice_to_openai({"type": "none"})
     assert result == "none"
+
+
+@pytest.mark.parametrize(
+    "model, expected",
+    [
+        ("anthropic/claude-sonnet-4", True),
+        ("claude-3-5-sonnet", True),
+        ("openrouter/qwen/qwen3-max", True),
+        # DeepSeek caching is automatic/prefix-based and ignores cache_control,
+        # so it is intentionally NOT recognized here (no cache_control applied).
+        ("openrouter/deepseek/deepseek-v3", False),
+        ("openrouter/openai/gpt-4o", False),
+        ("gemini/gemini-1.5-pro", False),
+    ],
+)
+def test_is_anthropic_claude_model_includes_qwen(model, expected):
+    """
+    cache_control passthrough in the Anthropic->OpenAI adapter is gated on
+    is_anthropic_claude_model. Alibaba (Qwen) upstreams require explicit
+    cache_control breakpoints on OpenRouter, so qwen must be recognized here;
+    DeepSeek must not (its caching is automatic and ignores cache_control).
+    """
+    assert LiteLLMAnthropicMessagesAdapter.is_anthropic_claude_model(model) is expected
+
+
+def test_add_system_message_filters_billing_header_block():
+    """
+    The `x-anthropic-billing-header:` system block carries a per-request hash
+    that invalidates the upstream prefix cache on every turn. It must be
+    dropped during the Anthropic->OpenAI system-block translation, mirroring
+    _filter_billing_headers_from_system in the messages adapter.
+    """
+    adapter = LiteLLMAnthropicMessagesAdapter()
+    new_messages: list = []
+    request = {
+        "model": "openrouter/qwen/qwen3-max",
+        "system": [
+            {"type": "text", "text": "x-anthropic-billing-header: cch=abcde"},
+            {
+                "type": "text",
+                "text": "You are a helpful assistant.",
+                "cache_control": {"type": "ephemeral"},
+            },
+        ],
+    }
+
+    adapter._add_system_message_to_messages(new_messages, cast(Any, request))
+
+    assert len(new_messages) == 1
+    content = new_messages[0]["content"]
+    texts = [block["text"] for block in content]
+    assert "You are a helpful assistant." in texts
+    assert all(not t.startswith("x-anthropic-billing-header:") for t in texts)
+    # cache_control survives for qwen (gated by is_anthropic_claude_model)
+    assert any(block.get("cache_control") for block in content)

tests/test_litellm/llms/openrouter/chat/test_openrouter_chat_transformation.py

@@ -553,3 +553,63 @@ def test_openrouter_non_reasoning_models_do_not_add_reasoning_effort():
     )
 
     assert "reasoning_effort" not in supported_params
+
+
+def test_openrouter_transform_request_with_cache_control_qwen():
+    """
+    Alibaba (Qwen) upstreams require explicit cache_control breakpoints on
+    OpenRouter, so transform_request must preserve the marker (move it to a
+    content block) rather than stripping it. Without this every turn pays
+    full input rate.
+    """
+    config = OpenrouterConfig()
+
+    messages = [
+        {
+            "role": "user",
+            "content": "Analyze this data",
+            "cache_control": {"type": "ephemeral"},
+        }
+    ]
+
+    transformed_request = config.transform_request(
+        model="openrouter/qwen/qwen3-max",
+        messages=messages,
+        optional_params={},
+        litellm_params={},
+        headers={},
+    )
+
+    user_message = transformed_request["messages"][0]
+    assert isinstance(user_message["content"], list)
+    assert user_message["content"][0]["type"] == "text"
+    assert user_message["content"][0]["cache_control"] == {"type": "ephemeral"}
+
+
+def test_openrouter_transform_request_drops_cache_control_for_deepseek():
+    """
+    DeepSeek caching on OpenRouter is automatic/prefix-based and ignores
+    cache_control, so DeepSeek is intentionally absent from
+    CacheControlSupportedModels. The handler should strip the marker rather
+    than emit an inert cache_control content block.
+    """
+    config = OpenrouterConfig()
+
+    messages = [
+        {
+            "role": "user",
+            "content": "Analyze this data",
+            "cache_control": {"type": "ephemeral"},
+        }
+    ]
+
+    transformed_request = config.transform_request(
+        model="openrouter/deepseek/deepseek-v3",
+        messages=messages,
+        optional_params={},
+        litellm_params={},
+        headers={},
+    )
+
+    user_message = transformed_request["messages"][0]
+    assert "cache_control" not in user_message