Fix issue #5383: [Bug]: LLM Cost is added to the metrics twice

openhands/llm/llm.py

@@ -219,6 +219,20 @@ def wrapper(*args, **kwargs):
                         )
                     resp.choices[0].message = fn_call_response_message
 
+                message_back: str = resp['choices'][0]['message']['content'] or ''
+                tool_calls = resp['choices'][0]['message'].get('tool_calls', [])
+                if tool_calls:
+                    for tool_call in tool_calls:
+                        fn_name = tool_call.function.name
+                        fn_args = tool_call.function.arguments
+                        message_back += f'\nFunction call: {fn_name}({fn_args})'
+
+                # log the LLM response
+                self.log_response(message_back)
+
+                # post-process the response first to calculate cost
+                self._post_completion(resp)
+
                 # log for evals or other scripts that need the raw completion
                 if self.config.log_completions:
                     assert self.config.log_completions_folder is not None
@@ -228,13 +242,15 @@ def wrapper(*args, **kwargs):
                         f'{self.metrics.model_name.replace("/", "__")}-{time.time()}.json',
                     )
 
+                    # Get cost from metrics instead of recalculating
+                    cost = self.metrics.costs[-1]['cost'] if self.metrics.costs else 0.0
                     _d = {
                         'messages': messages,
                         'response': resp,
                         'args': args,
                         'kwargs': {k: v for k, v in kwargs.items() if k != 'messages'},
                         'timestamp': time.time(),
-                        'cost': self._completion_cost(resp),
+                        'cost': cost,
                     }
                     if mock_function_calling:
                         # Overwrite response as non-fncall to be consistent with `messages``
@@ -245,20 +261,6 @@ def wrapper(*args, **kwargs):
                     with open(log_file, 'w') as f:
                         f.write(json.dumps(_d))
 
-                message_back: str = resp['choices'][0]['message']['content'] or ''
-                tool_calls = resp['choices'][0]['message'].get('tool_calls', [])
-                if tool_calls:
-                    for tool_call in tool_calls:
-                        fn_name = tool_call.function.name
-                        fn_args = tool_call.function.arguments
-                        message_back += f'\nFunction call: {fn_name}({fn_args})'
-
-                # log the LLM response
-                self.log_response(message_back)
-
-                # post-process the response
-                self._post_completion(resp)
-
                 return resp
             except APIError as e:
                 if 'Attention Required! | Cloudflare' in str(e):
@@ -534,7 +536,8 @@ def _completion_cost(self, response) -> float:
                 cost = litellm_completion_cost(
                     completion_response=response, **extra_kwargs
                 )
-            self.metrics.add_cost(cost)
+                # Only add cost to metrics if we're calculating it for the first time
+                self.metrics.add_cost(cost)
             return cost
         except Exception:
             self.cost_metric_supported = False