@@ -0,0 +1 @@
+Record cache-write and audio token usage from OpenAI Chat Completions responses and streams.
@@ -7,6 +7,7 @@
 import logging
 
 from openai import AsyncStream, Stream
+from openai.types import CompletionUsage
 from openai.types.chat import ChatCompletionChunk
 
 from opentelemetry.semconv._incubating.attributes import (
@@ -25,7 +26,10 @@
 )
 
 from .chat_buffers import ChoiceBuffer
-from .utils import get_property_value, map_finish_reason
+from .utils import (
+    map_finish_reason,
+    set_chat_usage,
+)
 
 _logger = logging.getLogger(__name__)
 
@@ -38,10 +42,7 @@ class _ChatStreamMixin:
     _self_choice_buffers: list[ChoiceBuffer]
     _self_response_id: str | None
     _self_service_tier: str | None
-    _self_prompt_tokens: int | None
-    _self_completion_tokens: int | None
-    _self_cached_prompt_tokens: int | None
-    _self_reasoning_tokens: int | None
+    _self_usage: CompletionUsage | None
 
     def _set_response_model(self, chunk: ChatCompletionChunk) -> None:
         # Set eagerly so the per-chunk streaming timing metrics carry
@@ -94,26 +95,6 @@ def _build_streaming_response(self, chunk: ChatCompletionChunk) -> None:
                         tool_call
                     )
 
-    def _set_usage(self, chunk: ChatCompletionChunk) -> None:
-        usage = getattr(chunk, "usage", None)
-        if usage:
-            self._self_completion_tokens = usage.completion_tokens
-            self._self_prompt_tokens = usage.prompt_tokens
-            prompt_tokens_details = getattr(
-                usage, "prompt_tokens_details", None
-            )
-            if prompt_tokens_details is not None:
-                self._self_cached_prompt_tokens = get_property_value(
-                    prompt_tokens_details, "cached_tokens"
-                )
-            completion_tokens_details = getattr(
-                usage, "completion_tokens_details", None
-            )
-            if completion_tokens_details is not None:
-                self._self_reasoning_tokens = get_property_value(
-                    completion_tokens_details, "reasoning_tokens"
-                )
-
     def _process_chunk(self, chunk: ChatCompletionChunk) -> None:
         if not isinstance(chunk, ChatCompletionChunk):
             # raw-response stream can be parsed into a caller-defined chunk type.
@@ -127,7 +108,9 @@ def _process_chunk(self, chunk: ChatCompletionChunk) -> None:
         self._set_response_model(chunk)
         self._set_response_service_tier(chunk)
         self._build_streaming_response(chunk)
-        self._set_usage(chunk)
+        usage: CompletionUsage | None = chunk.usage
+        if usage is not None:
+            self._self_usage = usage
 
     def _set_output_messages(self) -> None:
         if not self._self_capture_content:  # optimization
@@ -172,12 +155,10 @@ def _on_stream_error(self, error: BaseException) -> None:
 
     def _cleanup(self, error: BaseException | None = None) -> None:
         self._self_invocation.response_id = self._self_response_id
-        self._self_invocation.input_tokens = self._self_prompt_tokens
-        self._self_invocation.output_tokens = self._self_completion_tokens
-        self._self_invocation.cache_read_input_tokens = (
-            self._self_cached_prompt_tokens
-        )
-        self._self_invocation.thinking_tokens = self._self_reasoning_tokens
+        if self._self_usage is not None:
+            set_chat_usage(
+                invocation=self._self_invocation, usage=self._self_usage
+            )
         finish_reasons = [
             choice.finish_reason
             for choice in self._self_choice_buffers
@@ -217,10 +198,7 @@ def __init__(
         self._self_capture_content = capture_content
         self._self_response_id = None
         self._self_service_tier = None
-        self._self_prompt_tokens = None
-        self._self_completion_tokens = None
-        self._self_cached_prompt_tokens = None
-        self._self_reasoning_tokens = None
+        self._self_usage = None
 
 
 class AsyncChatStreamWrapper(
@@ -240,10 +218,7 @@ def __init__(
         self._self_capture_content = capture_content
         self._self_response_id = None
         self._self_service_tier = None
-        self._self_prompt_tokens = None
-        self._self_completion_tokens = None
-        self._self_cached_prompt_tokens = None
-        self._self_reasoning_tokens = None
+        self._self_usage = None
 
 
 __all__ = [
 
@@ -5,6 +5,7 @@
 import logging
 
 from openai.types import CreateEmbeddingResponse
+from openai.types.chat import ChatCompletion
 
 from opentelemetry.semconv._incubating.attributes import (
     gen_ai_attributes as GenAIAttributes,
@@ -23,10 +24,10 @@
 from .utils import (
     _prepare_output_messages,
     create_chat_invocation,
-    get_property_value,
     get_server_address_and_port,
     get_value,
     is_streaming,
+    set_chat_usage,
 )
 
 _logger = logging.getLogger(__name__)
@@ -158,7 +159,9 @@ async def traced_method(wrapped, instance, args, kwargs):
 
 
 def _set_response_properties(
-    chat_invocation: InferenceInvocation, result, capture_content: bool
+    chat_invocation: InferenceInvocation,
+    result: ChatCompletion | ParsableResponse,
+    capture_content: bool,
 ) -> InferenceInvocation:
     if isinstance(result, ParsableResponse):
         # with_raw_response: safe to parse() here since this is the
@@ -196,22 +199,7 @@ def _set_response_properties(
         )
 
     if getattr(result, "usage", None):
-        chat_invocation.input_tokens = result.usage.prompt_tokens
-        chat_invocation.output_tokens = result.usage.completion_tokens
-        prompt_tokens_details = getattr(
-            result.usage, "prompt_tokens_details", None
-        )
-        if prompt_tokens_details is not None:
-            chat_invocation.cache_read_input_tokens = get_property_value(
-                prompt_tokens_details, "cached_tokens"
-            )
-        completion_tokens_details = getattr(
-            result.usage, "completion_tokens_details", None
-        )
-        if completion_tokens_details is not None:
-            chat_invocation.thinking_tokens = get_property_value(
-                completion_tokens_details, "reasoning_tokens"
-            )
+        set_chat_usage(invocation=chat_invocation, usage=result.usage)
 
     if getattr(result, "system_fingerprint", None):
         chat_invocation.attributes.update(
 
@@ -11,6 +11,7 @@
 
 import openai
 from openai import NotGiven
+from openai.types import CompletionUsage
 
 from opentelemetry.semconv._incubating.attributes import (
     gen_ai_attributes as GenAIAttributes,
@@ -68,6 +69,50 @@ def get_property_value(obj, property_name):
     return getattr(obj, property_name, None)
 
 
+def set_chat_usage(
+    invocation: InferenceInvocation, usage: CompletionUsage
+) -> None:
+    invocation.input_tokens = usage.prompt_tokens
+    invocation.output_tokens = usage.completion_tokens
+    prompt_details: object = get_property_value(
+        obj=usage, property_name="prompt_tokens_details"
+    )
+    completion_details: object = get_property_value(
+        obj=usage, property_name="completion_tokens_details"
+    )
+    invocation.cache_read_input_tokens = get_property_value(
+        obj=prompt_details, property_name="cached_tokens"
+    )
+    invocation.cache_write_input_tokens = get_property_value(
+        obj=prompt_details, property_name="cache_write_tokens"
+    )
+    invocation.thinking_tokens = get_property_value(
+        obj=completion_details, property_name="reasoning_tokens"
+    )
+    invocation.set_input_tokens(
+        entries=(
+            (
+                "audio",
+                get_property_value(
+                    obj=prompt_details,
+                    property_name="audio_tokens",
+                ),
+            ),
+        )
+    )
+    invocation.set_output_tokens(
+        entries=(
+            (
+                "audio",
+                get_property_value(
+                    obj=completion_details,
+                    property_name="audio_tokens",
+                ),
+            ),
+        )
+    )
+
+
 def get_server_address_and_port(
     client_instance,
 ) -> tuple[str | None, int | None]:
Original file line number	Diff line number	Diff line change
`@@ -0,0 +1 @@`
	`1`	`+Record cache-write and audio token usage from OpenAI Chat Completions responses and streams.`