From 8d90603ba834b36d292da4fe626b05f431f6fa2f Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Thu, 5 Nov 2026 21:46:55 -0800
Subject: [PATCH] fix(cache): keep /v1/messages marks only where the chat route
 turns them into breakpoints

---
 .../anthropic/messages/adapters/handler.py    | 13 ++++-
 .../adapters/test_handler_prompt_cache_key.py | 52 +++++++++++++++++++
 2 files changed, 64 insertions(+), 1 deletion(-)

diff --git a/litellm/llms/anthropic/messages/adapters/handler.py b/litellm/llms/anthropic/messages/adapters/handler.py
index 329594c75d7..2d2c4680a95 100644
--- a/litellm/llms/anthropic/messages/adapters/handler.py
+++ b/litellm/llms/anthropic/messages/adapters/handler.py
@@ -397,9 +397,20 @@ ANTHROPIC_ADAPTER: Final = AnthropicAdapter()
 ########################################################
 
 
+_ROUTES_FORWARDING_CACHE_CONTROL_UNMAPPED: Final = frozenset(("azure",))
+_ROUTES_READING_ONLY_THE_SERVED_MODEL: Final = frozenset(("openai",))
+
+
 def _target_reads_prompt_cache_breakpoint(model: str, call_kwargs: Mapping[str, object]) -> bool:
+    provider: Final = call_kwargs.get("custom_llm_provider")
+    if provider in _ROUTES_FORWARDING_CACHE_CONTROL_UNMAPPED:
+        return False
     base_model: Final = call_kwargs.get("base_model")
-    named_models: Final = (model, base_model) if isinstance(base_model, str) and base_model else (model,)
+    named_models: Final = (
+        (model, base_model)
+        if isinstance(base_model, str) and base_model and provider not in _ROUTES_READING_ONLY_THE_SERVED_MODEL
+        else (model,)
+    )
     return any(call_targets_openai_prompt_cache_breakpoint(name, call_kwargs) for name in named_models)
 
 
diff --git a/tests/unit/llms/anthropic/messages/adapters/test_handler_prompt_cache_key.py b/tests/unit/llms/anthropic/messages/adapters/test_handler_prompt_cache_key.py
index 9d08632e29b..11753a49aea 100644
--- a/tests/unit/llms/anthropic/messages/adapters/test_handler_prompt_cache_key.py
+++ b/tests/unit/llms/anthropic/messages/adapters/test_handler_prompt_cache_key.py
@@ -224,3 +224,55 @@ def test_anthropic_messages_to_a_deployment_that_opts_out_of_breakpoints_carry_n
 
     assert list(_texts_with("cache_control", completion_kwargs)) == []
     assert list(_texts_with("prompt_cache_breakpoint", wire)) == []
+
+
+_OPENAI_GPT_ROW: Final = {"litellm_provider": "openai", "mode": "chat", "supports_prompt_cache_breakpoint": True}
+
+
+@pytest.mark.parametrize(
+    ("model", "extra_kwargs", "keeps_marks"),
+    [
+        pytest.param(
+            "azure/gpt-5.6-sol",
+            {"custom_llm_provider": "azure", "model_info": {"supports_prompt_cache_breakpoint": True}},
+            False,
+            id="azure-chat-forwards-cache-control-unmapped",
+        ),
+        pytest.param(
+            "openai/team-alias",
+            {
+                "custom_llm_provider": "openai",
+                "base_model": "gpt-5.6-sol",
+                "api_base": "https://llm.example.com/v1",
+                "prompt_cache_options": {"mode": "explicit"},
+            },
+            False,
+            id="openai-chat-reads-only-the-served-model",
+        ),
+        pytest.param(
+            "bedrock_mantle/team-alias",
+            {
+                "custom_llm_provider": "bedrock_mantle",
+                "base_model": "openai.gpt-5.6-sol",
+                "prompt_cache_options": {"mode": "explicit"},
+            },
+            True,
+            id="bridge-reads-base-model",
+        ),
+    ],
+)
+def test_anthropic_messages_keep_marks_only_for_a_route_that_turns_them_into_breakpoints(
+    monkeypatch: pytest.MonkeyPatch, model: str, extra_kwargs: dict[str, object], keeps_marks: bool
+) -> None:
+    monkeypatch.setitem(litellm.model_cost, "gpt-5.6-sol", _OPENAI_GPT_ROW)
+    monkeypatch.setitem(litellm.model_cost, "bedrock_mantle/openai.gpt-5.6-sol", _MANTLE_GPT_ROW)
+
+    completion_kwargs, _ = LiteLLMMessagesToCompletionTransformationHandler._prepare_completion_kwargs(
+        max_tokens=64,
+        messages=json.loads(json.dumps(_CLAUDE_CODE_MESSAGES)),
+        model=model,
+        system=json.loads(json.dumps(_SYSTEM)),
+        extra_kwargs=extra_kwargs,
+    )
+
+    assert bool(list(_texts_with("cache_control", completion_kwargs["messages"]))) is keeps_marks