fix: keep video fallback for usage without prompt tokens

litellm/cost_calculator.py

@@ -375,26 +375,28 @@ def _get_max_input_tokens_for_cost_fallback(
     return None
 
 
-def _usage_has_token_counts(usage_object: Optional[Usage]) -> bool:
+def _usage_has_prompt_token_counts(usage_object: Optional[Usage]) -> bool:
     if usage_object is None:
         return False
 
-    for attr in ("prompt_tokens", "completion_tokens", "total_tokens"):
-        if _is_positive_finite_number(getattr(usage_object, attr, 0)):
-            return True
+    if _is_positive_finite_number(getattr(usage_object, "prompt_tokens", 0)):
+        return True
 
     prompt_details = getattr(usage_object, "prompt_tokens_details", None)
     if prompt_details is not None:
-        for attr in ("audio_tokens", "cached_tokens", "text_tokens"):
+        for attr in (
+            "audio_tokens",
+            "cached_tokens",
+            "cache_creation_tokens",
+            "character_count",
+            "image_count",
+            "image_tokens",
+            "text_tokens",
+            "video_length_seconds",
+        ):
             if _is_positive_finite_number(getattr(prompt_details, attr, 0)):
                 return True
 
-    completion_details = getattr(usage_object, "completion_tokens_details", None)
-    if completion_details is not None:
-        for attr in ("audio_tokens", "reasoning_tokens", "text_tokens"):
-            if _is_positive_finite_number(getattr(completion_details, attr, 0)):
-                return True
-
     return False
 
 
@@ -1684,7 +1686,7 @@ def completion_cost(  # noqa: PLR0915
 
                 if (
                     call_type in _CHAT_COMPLETION_CALL_TYPES
-                    and not _usage_has_token_counts(cost_per_token_usage_object)
+                    and not _usage_has_prompt_token_counts(cost_per_token_usage_object)
                 ):
                     # Video metadata is client-provided. When provider usage is
                     # missing, keep spend reconciliation conservative.

tests/local_testing/test_completion_cost.py

@@ -1031,6 +1031,69 @@ def test_completion_cost_uses_conservative_video_fallback_without_usage():
     assert cost == pytest.approx(max_input_tokens * input_cost_per_token)
 
 
+@pytest.mark.parametrize(
+    "usage,expected_completion_tokens",
+    [
+        ({"total_tokens": 5}, 0),
+        ({"completion_tokens": 3, "total_tokens": 3}, 3),
+    ],
+)
+def test_completion_cost_uses_video_fallback_without_prompt_usage(
+    usage,
+    expected_completion_tokens,
+):
+    model = "openai/test-video-cost-fallback-no-prompt-usage"
+    input_cost_per_token = 0.25
+    output_cost_per_token = 0.5
+    max_input_tokens = 16
+    litellm.register_model(
+        model_cost={
+            model: {
+                "input_cost_per_token": input_cost_per_token,
+                "output_cost_per_token": output_cost_per_token,
+                "max_tokens": max_input_tokens,
+                "max_input_tokens": max_input_tokens,
+                "max_output_tokens": 4,
+                "litellm_provider": "openai",
+                "mode": "chat",
+            }
+        }
+    )
+    messages = [
+        {
+            "role": "user",
+            "content": [
+                {
+                    "type": "video_url",
+                    "video_url": {
+                        "url": "https://example.com/video.mp4",
+                        "video_metadata": {
+                            "duration_seconds": 0,
+                            "fps": 0,
+                            "has_audio": False,
+                        },
+                    },
+                }
+            ],
+        }
+    ]
+
+    try:
+        cost = completion_cost(
+            completion_response={"model": model, "usage": usage},
+            model=model,
+            messages=messages,
+            custom_llm_provider="openai",
+        )
+    finally:
+        litellm.model_cost.pop(model, None)
+
+    assert cost == pytest.approx(
+        (max_input_tokens * input_cost_per_token)
+        + (expected_completion_tokens * output_cost_per_token)
+    )
+
+
 @pytest.mark.parametrize("metadata_key", ["metadata", "litellm_metadata"])
 def test_completion_cost_ignores_client_metadata_for_video_fallback_limit(
     metadata_key,