diff --git a/langfun/core/llms/anthropic.py b/langfun/core/llms/anthropic.py
index 83498e2b..f9890299 100644
--- a/langfun/core/llms/anthropic.py
+++ b/langfun/core/llms/anthropic.py
@@ -54,8 +54,9 @@ class RateLimits(lf.ModelInfo.RateLimits):
 
     @property
     def max_tokens_per_minute(self) -> int:
-      return (self.max_input_tokens_per_minute
-              + self.max_output_tokens_per_minute)
+      return (
+          self.max_input_tokens_per_minute + self.max_output_tokens_per_minute
+      )
 
 
 SUPPORTED_MODELS = [
@@ -839,9 +840,7 @@ class Anthropic(rest.REST):
   """
 
   model: pg.typing.Annotated[
-      pg.typing.Enum(
-          pg.MISSING_VALUE, [m.model_id for m in SUPPORTED_MODELS]
-      ),
+      pg.typing.Enum(pg.MISSING_VALUE, [m.model_id for m in SUPPORTED_MODELS]),
       'The name of the model to use.',
   ]
 
@@ -855,10 +854,7 @@ class Anthropic(rest.REST):
 
   api_endpoint: str = 'https://api.anthropic.com/v1/messages'
 
-  api_version: Annotated[
-      str,
-      'Anthropic API version.'
-  ] = '2023-06-01'
+  api_version: Annotated[str, 'Anthropic API version.'] = '2023-06-01'
 
   thinking: Annotated[
       bool | None,
@@ -912,9 +908,7 @@ def _use_adaptive_thinking(self) -> bool:
     return self.model is not None and 'claude-opus-4-7' in self.model_id
 
   def request(
-      self,
-      prompt: lf.Message,
-      sampling_options: lf.LMSamplingOptions
+      self, prompt: lf.Message, sampling_options: lf.LMSamplingOptions
   ) -> dict[str, Any]:
     """Returns the JSON input for a message."""
     request = dict()
@@ -1022,14 +1016,27 @@ def _request_args(self, options: lf.LMSamplingOptions) -> dict[str, Any]:
 
   def result(self, json: dict[str, Any]) -> lf.LMSamplingResult:
     message = lf.Message.from_value(json, format='anthropic')
-    input_tokens = json['usage']['input_tokens']
-    output_tokens = json['usage']['output_tokens']
+    usage = json.get('usage', {})
+    input_tokens = usage.get('input_tokens', 0)
+    output_tokens = usage.get('output_tokens', 0)
+    cache_read_tokens = usage.get('cache_read_input_tokens', 0)
+    cache_creation_tokens = usage.get('cache_creation_input_tokens', 0)
+
+    # Anthropic's input_tokens excludes cache hits. Total prompt tokens
+    # comprises both cached and uncached segments.
+    prompt_tokens = input_tokens + cache_read_tokens + cache_creation_tokens
+
     return lf.LMSamplingResult(
         [lf.LMSample(message)],
         usage=lf.LMSamplingUsage(
-            prompt_tokens=input_tokens,
+            prompt_tokens=prompt_tokens,
             completion_tokens=output_tokens,
-            total_tokens=input_tokens + output_tokens,
+            total_tokens=prompt_tokens + output_tokens,
+            cached_prompt_tokens=cache_read_tokens,
+            completion_tokens_details={
+                'cache_creation_input_tokens': cache_creation_tokens,
+                'cache_read_input_tokens': cache_read_tokens,
+            },
         ),
     )
 
@@ -1118,21 +1125,25 @@ class Claude35(Anthropic):
 
 class Claude35Sonnet(Claude35):
   """Claude 3.5 Sonnet model (latest)."""
+
   model = 'claude-3-5-sonnet-latest'
 
 
 class Claude35Sonnet_20241022(Claude35):  # pylint: disable=invalid-name
   """Claude 3.5 Sonnet model (10/22/2024)."""
+
   model = 'claude-3-5-sonnet-20241022'
 
 
 class Claude35Haiku(Claude35):
   """Claude 3.5 Haiku model (latest)."""
+
   model = 'claude-3-5-haiku-latest'
 
 
 class Claude35Haiku_20241022(Claude35):  # pylint: disable=invalid-name
   """Claude 3.5 Haiku model (10/22/2024)."""
+
   model = 'claude-3-5-haiku-20241022'
 
 
@@ -1182,4 +1193,5 @@ def _register_anthropic_models():
     if m.provider == 'Anthropic':
       lf.LanguageModel.register(m.model_id, Anthropic)
 
+
 _register_anthropic_models()