Merge d7295c15a3 into 9071ca503e

litellm/cost_calculator.py

@@ -310,6 +310,17 @@ def _transcription_usage_has_token_details(
     return (prompt_tokens_val > 0) or (completion_tokens_val > 0)
 
 
+def _coerce_token_count(value: float | str | None) -> int:
+    if value is None:
+        return 0
+    if isinstance(value, str):
+        try:
+            return int(value)
+        except ValueError:
+            return 0
+    return int(value)
+
+
 def cost_per_token(
     model: str = "",
     prompt_tokens: int = 0,
@@ -367,14 +378,17 @@ def cost_per_token(
     if model is None:
         raise Exception("Invalid arg. Model cannot be none.")
 
+    prompt_token_count = _coerce_token_count(prompt_tokens)
+    completion_token_count = _coerce_token_count(completion_tokens)
+
     ## RECONSTRUCT USAGE BLOCK ##
     if usage_object is not None:
         usage_block = usage_object
     else:
         usage_block = Usage(
-            prompt_tokens=prompt_tokens,
-            completion_tokens=completion_tokens,
-            total_tokens=prompt_tokens + completion_tokens,
+            prompt_tokens=prompt_token_count,
+            completion_tokens=completion_token_count,
+            total_tokens=prompt_token_count + completion_token_count,
             cache_creation_input_tokens=cache_creation_input_tokens,
             cache_read_input_tokens=cache_read_input_tokens,
         )
@@ -418,13 +432,13 @@ def cost_per_token(
 
     # Anthropic reports prompt_tokens as input_tokens (excluding cache tokens).
     # Adjust so the helper's "prompt_tokens includes cache tokens" invariant holds.
-    _normalized_prompt_tokens = float(prompt_tokens)
+    _normalized_prompt_tokens = float(prompt_token_count)
     if _is_anthropic_style:
         _normalized_prompt_tokens += _cache_read_tokens + _cache_creation_tokens
 
     response_cost: Final = _cost_per_token_custom_pricing_helper(
         prompt_tokens=_normalized_prompt_tokens,
-        completion_tokens=completion_tokens,
+        completion_tokens=completion_token_count,
         response_time_ms=response_time_ms,
         cached_tokens=_cache_read_tokens,
         cache_creation_tokens=_cache_creation_tokens,

tests/test_litellm/test_cost_calculator.py

@@ -4768,3 +4768,30 @@ def test_collect_and_combine_realtime_usage_stores_partitioned_text_tokens() ->
     assert combined.completion_tokens_details.reasoning_tokens == 95
     assert combined.completion_tokens_details.text_tokens == 38
     assert combined.completion_tokens_details.audio_tokens == 0
+def test_cost_per_token_none_token_counts_cost_zero():
+    prompt_cost, completion_cost = cost_per_token(
+        model="gpt-4o-mini", prompt_tokens=None, completion_tokens=None
+    )
+    assert prompt_cost == 0.0
+    assert completion_cost == 0.0
+
+
+def test_cost_per_token_string_token_counts_coerced():
+    prompt_cost, completion_cost = cost_per_token(
+        model="gpt-4o-mini", prompt_tokens="10", completion_tokens=5
+    )
+    expected_prompt, expected_completion = cost_per_token(
+        model="gpt-4o-mini", prompt_tokens=10, completion_tokens=5
+    )
+    assert prompt_cost == expected_prompt
+    assert completion_cost == expected_completion
+
+
+def test_cost_per_token_garbage_string_counts_cost_zero():
+    prompt_cost, completion_cost = cost_per_token(
+        model="gpt-4o-mini", prompt_tokens="abc", completion_tokens=5
+    )
+    expected_prompt, _ = cost_per_token(
+        model="gpt-4o-mini", prompt_tokens=0, completion_tokens=5
+    )
+    assert prompt_cost == expected_prompt