diff --git a/src/social_hook/llm/litellm_client.py b/src/social_hook/llm/litellm_client.py
index be5bcf1..46b9dcb 100644
--- a/src/social_hook/llm/litellm_client.py
+++ b/src/social_hook/llm/litellm_client.py
@@ -116,10 +116,10 @@ def complete(
             # (token-field naming, reasoning params) instead of erroring.
             "drop_params": True,
         }
-        # The pipeline always passes exactly one tool — force it. (A provider
-        # that rejects forced tool_choice is caught by the JSON-from-text
-        # fallback in _extract_content.) "required" covers the rare multi-tool
-        # case.
+        # The pipeline always passes exactly one tool — force it. drop_params
+        # lets litellm drop a forced tool_choice the model can't honor; a model
+        # that then answers in prose is recovered by the JSON-from-text fallback
+        # in _extract_content. "required" covers the rare multi-tool case.
         if len(tools) == 1:
             kwargs["tool_choice"] = {"type": "function", "function": {"name": tools[0]["name"]}}
         elif tools:
@@ -183,7 +183,8 @@ def _extract_content(self, response: Any, tools: list[dict[str, Any]]) -> list[A
         if blocks:
             return blocks
 
-        text = (getattr(message, "content", None) or "").strip()
+        content = getattr(message, "content", None)
+        text = content.strip() if isinstance(content, str) else ""
         if len(tools) == 1 and text:
             try:
                 parsed = extract_json_object(text)
@@ -197,10 +198,14 @@ def _extract_content(self, response: Any, tools: list[dict[str, Any]]) -> list[A
 
     def _extract_usage(self, response: Any) -> NormalizedUsage:
         usage_obj = getattr(response, "usage", None)
-        input_tokens = int(getattr(usage_obj, "prompt_tokens", 0) or 0)
+        prompt_tokens = int(getattr(usage_obj, "prompt_tokens", 0) or 0)
         output_tokens = int(getattr(usage_obj, "completion_tokens", 0) or 0)
         ptd = getattr(usage_obj, "prompt_tokens_details", None)
         cache_read = int((getattr(ptd, "cached_tokens", 0) if ptd is not None else 0) or 0)
+        # OpenAI-style prompt_tokens INCLUDES cached tokens; normalize to the
+        # Anthropic-style disjoint accounting (non-cached input + separate
+        # cache_read) so cost and usage logging are consistent across providers.
+        input_tokens = max(prompt_tokens - cache_read, 0)
 
         cost_cents, cost_source = self._resolve_cost(
             response, input_tokens, output_tokens, cache_read
diff --git a/tests/test_litellm_client.py b/tests/test_litellm_client.py
index 5d1af67..9fc2903 100644
--- a/tests/test_litellm_client.py
+++ b/tests/test_litellm_client.py
@@ -148,6 +148,28 @@ def test_cost_falls_back_to_catalog(self, mock_completion, mock_cost):
         assert abs(resp.usage.cost_cents - 93.0) < 0.01
         assert resp.usage.cost_source == "registry"
 
+    @patch("litellm.completion_cost")
+    @patch("litellm.completion")
+    def test_cached_tokens_not_double_billed(self, mock_completion, mock_cost):
+        # OpenAI-style prompt_tokens INCLUDES cached tokens; the catalog
+        # fallback must bill the cached portion once (at 0.1x), not twice.
+        mock_completion.return_value = _mock_response(
+            tool_calls=[_tool_call("route_action", "{}")],
+            prompt_tokens=100_000,
+            cached_tokens=90_000,
+            completion_tokens=0,
+            response_cost=None,
+        )
+        mock_cost.side_effect = Exception("model not in litellm cost map")
+        resp = self._client().complete(messages=[{"role": "user", "content": "hi"}], tools=[TOOL])
+        # non-cached input = 10k; GLM-5.2 $0.93/M in, cache read at 0.1x:
+        # (10000*0.93 + 90000*0.93*0.1)/1e6*100 = 0.93 + 0.837 = 1.767c
+        assert abs(resp.usage.cost_cents - 1.767) < 0.01
+        assert resp.usage.cost_source == "registry"
+        # input_tokens is normalized to non-cached (Anthropic-style disjoint)
+        assert resp.usage.input_tokens == 10_000
+        assert resp.usage.cache_read_input_tokens == 90_000
+
     @patch("litellm.completion")
     def test_provider_error_wrapped(self, mock_completion):
         from social_hook.errors import MalformedResponseError