Refactor token accounting logic in utils.py

litellm/litellm_core_utils/llm_cost_calc/utils.py

@@ -668,9 +668,11 @@ def generic_cost_per_token(  # noqa: PLR0915
     image_tokens = prompt_tokens_details["image_tokens"]
 
     # Check for double-counting: sum of details > prompt_tokens means overlap
-    accounted_tokens = text_tokens + cache_hit + audio_tokens + cache_creation + image_tokens
+    accounted_tokens = (
+        text_tokens + cache_hit + audio_tokens + cache_creation + image_tokens
+    )
     has_double_counting = cache_hit > 0 and accounted_tokens > usage.prompt_tokens
- 
+
     # Some models use alternative billing dimensions (character_count, video_length_seconds)
     # that account for prompt_tokens without being included in the per-token detail fields.
     # When these are active, a gap between accounted_tokens and prompt_tokens is expected
@@ -679,7 +681,7 @@ def generic_cost_per_token(  # noqa: PLR0915
         prompt_tokens_details["character_count"] > 0
         or prompt_tokens_details["video_length_seconds"] > 0
     )
- 
+
     if has_double_counting:
         # Double-counting fix (xAI etc.): recalculate text_tokens from scratch
         # Clamp to 0 to prevent negative cost when cache_hit exceeds prompt_tokens
@@ -692,7 +694,11 @@ def generic_cost_per_token(  # noqa: PLR0915
             - image_tokens,
         )
         prompt_tokens_details["text_tokens"] = text_tokens
-    elif accounted_tokens < usage.prompt_tokens and not has_non_token_alternative_billing and prompt_tokens_details["image_count"] == 0:
+    elif (
+        accounted_tokens < usage.prompt_tokens
+        and not has_non_token_alternative_billing
+        and prompt_tokens_details["image_count"] == 0
+    ):
         # Unaccounted tokens fix: inline documents (PDF, DOCX, etc.) are counted
         # in prompt_tokens by the provider but not broken out into any detail field.
         # Add the gap to text_tokens so they are costed at input_cost_per_token.
@@ -707,8 +713,7 @@ def generic_cost_per_token(  # noqa: PLR0915
         #   billed separately via input_cost_per_image and shouldn't be double-charged
         unaccounted_tokens = usage.prompt_tokens - accounted_tokens
         prompt_tokens_details["text_tokens"] += unaccounted_tokens
- 
- 
+
     (
         prompt_base_cost,
         completion_base_cost,
@@ -718,7 +723,7 @@ def generic_cost_per_token(  # noqa: PLR0915
     ) = _get_token_base_cost(
         model_info=model_info, usage=usage, service_tier=service_tier
     )
- 
+
     prompt_cost = _calculate_input_cost(
         prompt_tokens_details=prompt_tokens_details,
         model_info=model_info,