diff --git a/backend/app/services/llm_service.py b/backend/app/services/llm_service.py
index 12d58931..01014486 100755
--- a/backend/app/services/llm_service.py
+++ b/backend/app/services/llm_service.py
@@ -188,19 +188,25 @@ class LLMService:
     @staticmethod
     def _extract_usage_metadata(response, provider: str) -> dict:
         """Extract token counts from a provider response. All fields default to 0."""
+        _log = logging.getLogger(__name__)
         if provider == 'gemini':
             um = getattr(response, 'usage_metadata', None)
             if um is None:
+                _log.warning("Gemini response missing usage_metadata — token counts will be 0, cost recorded as $0")
                 return {'prompt': 0, 'completion': 0, 'cached': 0, 'reasoning': 0}
+            # thoughts_token_count (thinking models) is already included in candidates_token_count.
+            # Capture it separately so the stored event can show the split.
+            thoughts = getattr(um, 'thoughts_token_count', 0) or 0
             return {
                 'prompt': getattr(um, 'prompt_token_count', 0) or 0,
                 'completion': getattr(um, 'candidates_token_count', 0) or 0,
                 'cached': getattr(um, 'cached_content_token_count', 0) or 0,
-                'reasoning': 0,
+                'reasoning': thoughts,
             }
         elif provider == 'openai':
             usage = getattr(response, 'usage', None)
             if usage is None:
+                _log.warning("OpenAI response missing usage — token counts will be 0, cost recorded as $0")
                 return {'prompt': 0, 'completion': 0, 'cached': 0, 'reasoning': 0}
             # Responses API (gpt-5.4-2026-03-05)
             if hasattr(usage, 'input_tokens'):