diff --git a/src/rius/generation.py b/src/rius/generation.py
index 61c9bf9..be1bdc6 100644
--- a/src/rius/generation.py
+++ b/src/rius/generation.py
@@ -27,6 +27,8 @@
     GEN_AI_REQUEST_PREFIX,
     GEN_AI_RESPONSE_FINISH_REASONS,
     GEN_AI_RESPONSE_MODEL,
+    GEN_AI_USAGE_CACHE_CREATION_INPUT_TOKENS,
+    GEN_AI_USAGE_CACHE_READ_INPUT_TOKENS,
     GEN_AI_USAGE_INPUT_TOKENS,
     GEN_AI_USAGE_OUTPUT_TOKENS,
     TRACER_NAME,
@@ -151,21 +153,46 @@ def set_response_model(self, response_model: str) -> None:
         self._span.set_attribute(GEN_AI_RESPONSE_MODEL, response_model)
 
     def set_usage(
-        self, *, input_tokens: int | None = None, output_tokens: int | None = None
+        self,
+        *,
+        input_tokens: int | None = None,
+        output_tokens: int | None = None,
+        cache_read_input_tokens: int | None = None,
+        cache_creation_input_tokens: int | None = None,
     ) -> None:
-        """Record token usage (``gen_ai.usage.input_tokens`` / ``output_tokens``).
+        """Record token usage (``gen_ai.usage.*`` attributes).
 
         Send token counts, never cost: cost is computed server-side from
         model pricing.
 
+        Pass provider-reported values as-is. Per the GenAI conventions,
+        ``input_tokens`` is the total including cached tokens (the cache
+        counts are subsets of it); some providers instead report an
+        ``input_tokens`` that excludes cache tokens (e.g. Anthropic, whose
+        OpenAI-style counterpart already includes them). The backend detects
+        and normalizes the exclusive case, so no client-side arithmetic is
+        needed.
+
         Args:
             input_tokens: Prompt tokens consumed, when known.
             output_tokens: Completion tokens produced, when known.
+            cache_read_input_tokens: Input tokens served from a
+                provider-managed prompt cache
+                (``gen_ai.usage.cache_read.input_tokens``).
+            cache_creation_input_tokens: Input tokens written to a
+                provider-managed prompt cache
+                (``gen_ai.usage.cache_creation.input_tokens``).
         """
         if input_tokens is not None:
             self._span.set_attribute(GEN_AI_USAGE_INPUT_TOKENS, input_tokens)
         if output_tokens is not None:
             self._span.set_attribute(GEN_AI_USAGE_OUTPUT_TOKENS, output_tokens)
+        if cache_read_input_tokens is not None:
+            self._span.set_attribute(GEN_AI_USAGE_CACHE_READ_INPUT_TOKENS, cache_read_input_tokens)
+        if cache_creation_input_tokens is not None:
+            self._span.set_attribute(
+                GEN_AI_USAGE_CACHE_CREATION_INPUT_TOKENS, cache_creation_input_tokens
+            )
 
     def record_first_token(self) -> None:
         """Mark the arrival of the first streamed token (``gen_ai.first_token`` event).
diff --git a/src/rius/semconv.py b/src/rius/semconv.py
index 7be0df0..112c683 100644
--- a/src/rius/semconv.py
+++ b/src/rius/semconv.py
@@ -40,6 +40,8 @@
 GEN_AI_RESPONSE_MODEL = "gen_ai.response.model"
 GEN_AI_USAGE_INPUT_TOKENS = "gen_ai.usage.input_tokens"
 GEN_AI_USAGE_OUTPUT_TOKENS = "gen_ai.usage.output_tokens"
+GEN_AI_USAGE_CACHE_READ_INPUT_TOKENS = "gen_ai.usage.cache_read.input_tokens"
+GEN_AI_USAGE_CACHE_CREATION_INPUT_TOKENS = "gen_ai.usage.cache_creation.input_tokens"
 GEN_AI_INPUT_MESSAGES = "gen_ai.input.messages"
 GEN_AI_OUTPUT_MESSAGES = "gen_ai.output.messages"
 GEN_AI_RESPONSE_FINISH_REASONS = "gen_ai.response.finish_reasons"
diff --git a/tests/test_generation.py b/tests/test_generation.py
index db46132..8ed0dab 100644
--- a/tests/test_generation.py
+++ b/tests/test_generation.py
@@ -52,6 +52,35 @@ def test_cm_output_and_usage(exported_spans: InMemorySpanExporter) -> None:
     assert attrs["gen_ai.response.model"] == "gpt-4o-2026-05"
 
 
+def test_cm_usage_cache_tokens(exported_spans: InMemorySpanExporter) -> None:
+    with start_as_current_generation("chat") as gen:
+        gen.set_usage(
+            input_tokens=10,
+            output_tokens=202,
+            cache_read_input_tokens=11579,
+            cache_creation_input_tokens=12694,
+        )
+    attrs = exported_spans.get_finished_spans()[0].attributes
+    assert attrs["gen_ai.usage.cache_read.input_tokens"] == 11579
+    assert attrs["gen_ai.usage.cache_creation.input_tokens"] == 12694
+
+
+def test_cm_usage_cache_tokens_omitted_are_absent(exported_spans: InMemorySpanExporter) -> None:
+    with start_as_current_generation("chat") as gen:
+        gen.set_usage(input_tokens=10, output_tokens=5)
+    attrs = exported_spans.get_finished_spans()[0].attributes
+    assert "gen_ai.usage.cache_read.input_tokens" not in attrs
+    assert "gen_ai.usage.cache_creation.input_tokens" not in attrs
+
+
+def test_cm_usage_cache_tokens_zero_recorded(exported_spans: InMemorySpanExporter) -> None:
+    with start_as_current_generation("chat") as gen:
+        gen.set_usage(cache_read_input_tokens=0, cache_creation_input_tokens=0)
+    attrs = exported_spans.get_finished_spans()[0].attributes
+    assert attrs["gen_ai.usage.cache_read.input_tokens"] == 0
+    assert attrs["gen_ai.usage.cache_creation.input_tokens"] == 0
+
+
 def test_cm_finish_reasons_list(exported_spans: InMemorySpanExporter) -> None:
     with start_as_current_generation("chat") as gen:
         gen.set_finish_reasons(["stop", "length"])