fix(anthropic): record reasoning token usage

packages/opentelemetry-instrumentation-anthropic/opentelemetry/instrumentation/anthropic/__init__.py

@@ -29,6 +29,7 @@
     count_prompt_tokens_from_request,
     dont_throw,
     error_metrics_attributes,
+    get_reasoning_tokens,
     run_async,
     set_span_attribute,
     shared_metrics_attributes,
@@ -197,6 +198,7 @@ async def _aset_token_usage(
     token_histogram: Histogram = None,
     choice_counter: Counter = None,
 ):
+    """Record token usage from an asynchronous Anthropic response."""
     import inspect
 
     # If we get a coroutine, await it
@@ -296,6 +298,12 @@ async def _aset_token_usage(
     )
     set_span_attribute(span, SpanAttributes.GEN_AI_USAGE_TOTAL_TOKENS, total_tokens)
 
+    set_span_attribute(
+        span,
+        SpanAttributes.GEN_AI_USAGE_REASONING_TOKENS,
+        get_reasoning_tokens(usage),
+    )
+
     set_span_attribute(
         span, GenAIAttributes.GEN_AI_USAGE_CACHE_READ_INPUT_TOKENS, cache_read_tokens
     )
@@ -316,6 +324,7 @@ def _set_token_usage(
     token_histogram: Histogram = None,
     choice_counter: Counter = None,
 ):
+    """Record token usage from a synchronous Anthropic response."""
     import inspect
 
     # If we get a coroutine, we cannot process it in sync context
@@ -411,7 +420,11 @@ def _set_token_usage(
         span, GenAIAttributes.GEN_AI_USAGE_OUTPUT_TOKENS, completion_tokens
     )
     set_span_attribute(span, SpanAttributes.GEN_AI_USAGE_TOTAL_TOKENS, total_tokens)
-
+    set_span_attribute(
+        span,
+        SpanAttributes.GEN_AI_USAGE_REASONING_TOKENS,
+        get_reasoning_tokens(usage),
+    )
     set_span_attribute(
         span, GenAIAttributes.GEN_AI_USAGE_CACHE_READ_INPUT_TOKENS, cache_read_tokens
     )

packages/opentelemetry-instrumentation-anthropic/opentelemetry/instrumentation/anthropic/streaming.py

@@ -15,6 +15,7 @@
     count_prompt_tokens_from_request,
     dont_throw,
     error_metrics_attributes,
+    get_reasoning_tokens,
     set_span_attribute,
     shared_metrics_attributes,
     should_emit_events,
@@ -82,6 +83,7 @@ def _set_token_usage(
     token_histogram: Histogram = None,
     choice_counter: Counter = None,
 ):
+    """Record token usage collected from a completed streaming response."""
     cache_read_tokens = (
         complete_response.get("usage", {}).get("cache_read_input_tokens", 0) or 0
     )
@@ -98,6 +100,12 @@ def _set_token_usage(
     )
     set_span_attribute(span, SpanAttributes.GEN_AI_USAGE_TOTAL_TOKENS, total_tokens)
 
+    set_span_attribute(
+        span,
+        SpanAttributes.GEN_AI_USAGE_REASONING_TOKENS,
+        get_reasoning_tokens(complete_response.get("usage")),
+    )
+
     set_span_attribute(
         span, GenAIAttributes.GEN_AI_USAGE_CACHE_READ_INPUT_TOKENS, cache_read_tokens
     )

packages/opentelemetry-instrumentation-anthropic/opentelemetry/instrumentation/anthropic/utils.py

@@ -27,6 +27,21 @@ def set_span_attribute(span, name, value):
     return
 
 
+def get_reasoning_tokens(usage):
+    """Return reasoning tokens from an Anthropic usage object or dictionary."""
+    if not usage:
+        return None
+
+    details = (
+        usage.get("output_tokens_details")
+        if isinstance(usage, dict)
+        else getattr(usage, "output_tokens_details", None)
+    )
+    if isinstance(details, dict):
+        return details.get("reasoning_tokens")
+    return getattr(details, "reasoning_tokens", None)
+
+
 def should_send_prompts():
     return (
         os.getenv(TRACELOOP_TRACE_CONTENT) or "true"

packages/opentelemetry-instrumentation-anthropic/tests/test_thinking.py

@@ -1,5 +1,12 @@
 import json
+from types import SimpleNamespace
+from unittest.mock import Mock
+
 import pytest
+from opentelemetry.instrumentation.anthropic import _set_token_usage
+from opentelemetry.instrumentation.anthropic.streaming import (
+    _set_token_usage as _set_stream_token_usage,
+)
 from opentelemetry.sdk._logs import ReadableLogRecord
 from opentelemetry.semconv._incubating.attributes import (
     gen_ai_attributes as GenAIAttributes,
@@ -8,6 +15,46 @@
 from .utils import verify_metrics
 
 
+def test_reasoning_tokens_are_recorded_for_non_streaming_response():
+    """Verify reasoning tokens are recorded for a regular response."""
+    span = Mock()
+    response = SimpleNamespace(
+        usage=SimpleNamespace(
+            input_tokens=10,
+            output_tokens=20,
+            output_tokens_details=SimpleNamespace(reasoning_tokens=7),
+        ),
+        content=[],
+        completion=None,
+        stop_reason=None,
+    )
+
+    _set_token_usage(span, None, {}, response)
+
+    span.set_attribute.assert_any_call(
+        GenAIAttributes.GEN_AI_USAGE_REASONING_TOKENS, 7
+    )
+
+
+def test_reasoning_tokens_are_recorded_for_streaming_response():
+    """Verify reasoning tokens are recorded for a streaming response."""
+    span = Mock()
+    complete_response = {
+        "usage": {
+            "input_tokens": 10,
+            "output_tokens": 20,
+            "output_tokens_details": {"reasoning_tokens": 7},
+        },
+        "events": [],
+    }
+
+    _set_stream_token_usage(span, complete_response, 10, 20)
+
+    span.set_attribute.assert_any_call(
+        GenAIAttributes.GEN_AI_USAGE_REASONING_TOKENS, 7
+    )
+
+
 @pytest.mark.vcr
 def test_anthropic_thinking_legacy(
     instrument_legacy, anthropic_client, span_exporter, log_exporter, reader