Token tweak

gemini/gemini.py

@@ -312,7 +312,6 @@ def main():
                 parts.append({"text": msg["text"]})
                 api_contents.append({"role": msg["role"], "parts": parts})
 
-            # Add the current prompt
             current_parts = []
             if not files_added_to_payload and not active_cache_id and file_objects:
                 for f in file_objects:
@@ -379,14 +378,13 @@ def main():
                 output_tokens = usage_metadata.candidates_token_count or 0
                 cached_tokens = getattr(usage_metadata, 'cached_content_token_count', 0) or 0
                 
-                # Handle API variations where prompt_token_count might strictly be the uncached tokens
-                if prompt_tokens >= cached_tokens:
-                    uncached_tokens = prompt_tokens - cached_tokens
-                else:
-                    uncached_tokens = prompt_tokens
-                    
-                total_input_tokens = uncached_tokens + cached_tokens
+                # Revert to max(0) to handle Google's padding discrepancy where prompt_tokens < cached_tokens
+                uncached_tokens = max(0, prompt_tokens - cached_tokens)
                 
+                # Ensure the tier logic checks the absolute largest representation of the payload
+                total_input_tokens = max(prompt_tokens, cached_tokens)
+                
+                # Fetch the rate dynamically from the parsed JSON or default to 0.0 if fetch failed
                 rates = pricing_data.get(args.model, {"input": 0.0, "cached": 0.0, "output": 0.0})
                 
                 input_rate = rates.get("input", 0.0)