diff --git a/agent/agent_runtime_helpers.py b/agent/agent_runtime_helpers.py
index d7d5c39ebd427..f4a5bd46c2f08 100644
--- a/agent/agent_runtime_helpers.py
+++ b/agent/agent_runtime_helpers.py
@@ -2092,6 +2092,13 @@ def anthropic_prompt_cache_policy(
     gateway implements the Anthropic cache_control contract
     (MiniMax, Zhipu GLM, LiteLLM's Anthropic proxy mode all do).
 
+    LiteLLM proxies exposing the OpenAI-compatible surface instead
+    (``/v1/chat/completions`` — e.g. ``provider: custom:litellm``)
+    also honour Anthropic-style ``cache_control`` markers when serving
+    a Claude model, and get the envelope layout (native_anthropic=False)
+    like OpenRouter.  Without this grant they serve zero cache hits,
+    re-billing the full prompt on every turn.
+
     Qwen / Alibaba-family models on OpenCode, OpenCode Go, and direct
     Alibaba (DashScope) also honour Anthropic-style ``cache_control``
     markers on OpenAI-wire chat completions. Upstream pi-mono #3392 /
@@ -2240,6 +2247,31 @@ def anthropic_prompt_cache_policy(
     if is_anthropic_wire and is_minimax_route:
         return True, True
 
+    # LiteLLM proxies exposing the OpenAI-compatible surface (e.g.
+    # ``provider: custom:litellm`` with /v1/chat/completions — /v1/messages
+    # returns 404) serve Claude models and honour Anthropic-style
+    # cache_control markers on the OpenAI wire, exactly like OpenRouter.
+    # Without this branch the OpenAI-wire LiteLLM route matches no grant
+    # branch and falls through to (False, False) — zero cache hits, the
+    # full prompt re-billed on every turn (the same silent failure class
+    # the Qwen branch below addresses). LiteLLM's Anthropic-native route
+    # (/v1/messages) is already covered by the third-party-gateway branch
+    # above, which returns the native layout.  Detection mirrors the
+    # MiniMax provider-or-host pattern: a ``litellm`` substring in the
+    # provider id (``custom:litellm``, ``litellm``) or in the base URL
+    # host (self-hosted proxies registered as bare ``custom``).  The host
+    # check is host-only via ``base_url_hostname`` so a ``litellm`` path
+    # segment on an unrelated host (e.g. ``https://api.corp.com/v1/litellm``)
+    # does not grant cache_control to a provider that may reject the
+    # marker with HTTP 400.
+    litellm_host = base_url_hostname(eff_base_url)
+    is_litellm = (
+        "litellm" in provider_lower
+        or "litellm" in litellm_host
+    )
+    if is_litellm and is_claude and not is_anthropic_wire:
+        return True, False
+
     # Qwen/Alibaba on OpenCode (Zen/Go) and native DashScope: OpenAI-wire
     # transport that accepts Anthropic-style cache_control markers and
     # rewards them with real cache hits.  Without this branch
diff --git a/tests/run_agent/test_anthropic_prompt_cache_policy.py b/tests/run_agent/test_anthropic_prompt_cache_policy.py
index e0a4ae00fa407..3bcfefe9548ae 100644
--- a/tests/run_agent/test_anthropic_prompt_cache_policy.py
+++ b/tests/run_agent/test_anthropic_prompt_cache_policy.py
@@ -312,6 +312,86 @@ def test_custom_openai_wire_does_not_cache_even_with_claude_name(self):
         assert agent._anthropic_prompt_cache_policy() == (False, False)
 
 
+class TestLiteLLMOpenAIWire:
+    """LiteLLM proxies exposing the OpenAI-compatible surface get cache_control.
+
+    A LiteLLM deployment that serves Claude over /v1/chat/completions
+    (``provider: custom:litellm``; /v1/messages returns 404) honours
+    Anthropic-style ``cache_control`` markers on the OpenAI wire. Before
+    this fix the policy matched no grant branch for this route and fell
+    through to (False, False) — zero cache hits, the full prompt re-billed
+    on every turn, with no error or warning. Envelope layout (native=False)
+    because the wire format is OpenAI chat.completions, like OpenRouter.
+    """
+
+    def test_named_litellm_custom_provider_caches_with_envelope_layout(self):
+        agent = _make_agent(
+            provider="custom:litellm",
+            base_url="https://litellm.example.com/v1",
+            api_mode="chat_completions",
+            model="claude-opus-5",
+        )
+        should, native = agent._anthropic_prompt_cache_policy()
+        assert should is True, "LiteLLM OpenAI-wire Claude must cache"
+        assert native is False, "LiteLLM OpenAI wire uses the envelope layout"
+
+    def test_bare_litellm_provider_caches(self):
+        agent = _make_agent(
+            provider="litellm",
+            base_url="https://litellm.example.com/v1",
+            api_mode="chat_completions",
+            model="claude-sonnet-4-6",
+        )
+        assert agent._anthropic_prompt_cache_policy() == (True, False)
+
+    def test_custom_provider_pointed_at_litellm_host_caches(self):
+        # A bare custom provider registered at a LiteLLM host — host match
+        # alone is sufficient, mirroring the MiniMax provider-or-host branch.
+        agent = _make_agent(
+            provider="custom",
+            base_url="https://litellm.example.com/v1",
+            api_mode="chat_completions",
+            model="claude-sonnet-4-6",
+        )
+        assert agent._anthropic_prompt_cache_policy() == (True, False)
+
+    def test_litellm_anthropic_wire_keeps_native_layout(self):
+        # LiteLLM's Anthropic-native route (/v1/messages) is covered by the
+        # third-party-gateway branch: native layout, not envelope.
+        agent = _make_agent(
+            provider="custom:litellm",
+            base_url="https://litellm.example.com/v1",
+            api_mode="anthropic_messages",
+            model="claude-opus-5",
+        )
+        assert agent._anthropic_prompt_cache_policy() == (True, True)
+
+    def test_non_claude_on_litellm_does_not_cache(self):
+        # LiteLLM serving a non-Claude model stays off: we only grant
+        # cache_control when the served model family is documented to
+        # honour the markers.
+        agent = _make_agent(
+            provider="custom:litellm",
+            base_url="https://litellm.example.com/v1",
+            api_mode="chat_completions",
+            model="openai/gpt-5.4",
+        )
+        assert agent._anthropic_prompt_cache_policy() == (False, False)
+
+    def test_litellm_path_segment_on_unrelated_host_does_not_cache(self):
+        # Host-only detection: a ``litellm`` substring confined to the URL
+        # path (e.g. a failover path on an unrelated provider) must NOT grant
+        # cache_control. Injecting the marker onto a strict OpenAI-wire host
+        # that rejects unknown keys would fail with HTTP 400 (#84506).
+        agent = _make_agent(
+            provider="custom",
+            base_url="https://api.corp.example.com/v1/litellm-failover",
+            api_mode="chat_completions",
+            model="claude-sonnet-4-6",
+        )
+        assert agent._anthropic_prompt_cache_policy() == (False, False)
+
+
 class TestQwenAlibabaFamily:
     """Qwen on OpenCode/OpenCode-Go/Alibaba — needs cache_control even on OpenAI-wire.