diff --git a/src/askui/models/openai/messages_api.py b/src/askui/models/openai/messages_api.py
index 83d8c4d4..ae023d90 100644
--- a/src/askui/models/openai/messages_api.py
+++ b/src/askui/models/openai/messages_api.py
@@ -291,8 +291,15 @@ def _from_openai_response(response: ChatCompletion) -> MessageParam:
         cached_tokens: int | None = None
         if response.usage.prompt_tokens_details is not None:
             cached_tokens = response.usage.prompt_tokens_details.cached_tokens
+        # OpenAI-style usage counts cached tokens INSIDE prompt_tokens, while
+        # `UsageParam` consumers (statistics callback, reporting) expect
+        # Anthropic-style disjoint fields — subtract so cached tokens are
+        # never counted or billed twice.
+        input_tokens = response.usage.prompt_tokens
+        if isinstance(cached_tokens, int) and cached_tokens > 0:
+            input_tokens = max(0, input_tokens - cached_tokens)
         usage = UsageParam(
-            input_tokens=response.usage.prompt_tokens,
+            input_tokens=input_tokens,
             output_tokens=response.usage.completion_tokens,
             cache_read_input_tokens=cached_tokens,
         )
diff --git a/tests/unit/models/openai/test_messages_api.py b/tests/unit/models/openai/test_messages_api.py
index b3ebfeb2..fe8b74c0 100644
--- a/tests/unit/models/openai/test_messages_api.py
+++ b/tests/unit/models/openai/test_messages_api.py
@@ -41,8 +41,14 @@ def _make_completion(
     finish_reason: str = "stop",
     prompt_tokens: int = 10,
     completion_tokens: int = 20,
+    cached_tokens: int | None = None,
 ) -> ChatCompletion:
     """Create a mock ChatCompletion response."""
+    prompt_tokens_details = None
+    if cached_tokens is not None:
+        from openai.types.completion_usage import PromptTokensDetails
+
+        prompt_tokens_details = PromptTokensDetails(cached_tokens=cached_tokens)
     return ChatCompletion(
         id="chatcmpl-test",
         choices=[
@@ -63,6 +69,7 @@ def _make_completion(
             prompt_tokens=prompt_tokens,
             completion_tokens=completion_tokens,
             total_tokens=prompt_tokens + completion_tokens,
+            prompt_tokens_details=prompt_tokens_details,
         ),
     )
 
@@ -456,6 +463,31 @@ def test_usage_captured(self) -> None:
         assert result.usage.input_tokens == 50
         assert result.usage.output_tokens == 100
 
+    def test_cached_tokens_are_subtracted_from_input(self) -> None:
+        """OpenAI reports cached tokens as a SUBSET of prompt_tokens;
+        `UsageParam` consumers expect disjoint fields (Anthropic style).
+        Without the subtraction, cached tokens are counted and billed twice
+        (statistics callback, reporting)."""
+        completion = _make_completion(
+            content="ok",
+            prompt_tokens=1000,
+            completion_tokens=50,
+            cached_tokens=400,
+        )
+        result = _from_openai_response(completion)
+        assert result.usage is not None
+        assert result.usage.input_tokens == 600  # 1000 - 400 cached
+        assert result.usage.cache_read_input_tokens == 400
+        assert result.usage.output_tokens == 50
+
+    def test_cached_tokens_never_drive_input_negative(self) -> None:
+        completion = _make_completion(
+            content="ok", prompt_tokens=100, completion_tokens=5, cached_tokens=150
+        )
+        result = _from_openai_response(completion)
+        assert result.usage is not None
+        assert result.usage.input_tokens == 0
+
 
 class TestOpenAIMessagesApi:
     def test_create_message_delegates_to_client(self) -> None: