Merge 68ca902ba9 into e26a6450c8

litellm/litellm_core_utils/streaming_chunk_builder_utils.py

@@ -1086,7 +1086,7 @@ class ChunkProcessor:
         except Exception:  # don't allow this failing to block a complete streaming response from being returned
             print_verbose("token_counter failed, assuming prompt tokens is 0")
             returned_usage.prompt_tokens = 0
-        returned_usage.completion_tokens = (
+        resolved_completion_tokens: Final = (
             completion_tokens
             if completion_tokens is not None
             else (
@@ -1097,7 +1097,9 @@ class ChunkProcessor:
                 )
                 + (reasoning_tokens or 0)
             )
+            + (reasoning_tokens or 0)
         )
+        returned_usage.completion_tokens = resolved_completion_tokens
         returned_usage.total_tokens = returned_usage.prompt_tokens + returned_usage.completion_tokens
 
         if cache_creation_input_tokens is not None:

tests/test_litellm/litellm_core_utils/test_streaming_chunk_builder_utils.py

@@ -1493,6 +1493,43 @@ def test_count_reasoning_tokens_counts_visible_reasoning():
     assert processor.count_reasoning_tokens(response) > 0
 
 
+def test_stream_chunk_builder_includes_reasoning_in_estimated_completion_tokens():
+    reasoning_content = "let me count the primes under thirty"
+    visible_content = "10"
+    chunks = [
+        ModelResponseStream(
+            id="chatcmpl-estimated-reasoning",
+            model="claude-opus-4-8",
+            choices=[
+                StreamingChoices(
+                    finish_reason=None,
+                    index=0,
+                    delta=Delta(role="assistant", reasoning_content=reasoning_content),
+                )
+            ],
+        ),
+        ModelResponseStream(
+            id="chatcmpl-estimated-reasoning",
+            model="claude-opus-4-8",
+            choices=[
+                StreamingChoices(
+                    finish_reason="stop",
+                    index=0,
+                    delta=Delta(content=visible_content),
+                )
+            ],
+        ),
+    ]
+
+    response = stream_chunk_builder(chunks=chunks, messages=[{"role": "user", "content": "count"}])
+
+    assert response is not None
+    reasoning_tokens = response.usage.completion_tokens_details.reasoning_tokens
+    assert reasoning_tokens is not None
+    assert response.usage.completion_tokens > reasoning_tokens
+    assert response.usage.total_tokens == response.usage.prompt_tokens + response.usage.completion_tokens
+
+
 @pytest.mark.parametrize(
     "estimated_reasoning_tokens, expected_reasoning_tokens, expected_text_tokens",
     [(40, 40, 60), (250, 100, 0)],