diff --git a/src/backend/crew_ai/crew.py b/src/backend/crew_ai/crew.py
index 702c83d..489f328 100644
--- a/src/backend/crew_ai/crew.py
+++ b/src/backend/crew_ai/crew.py
@@ -252,6 +252,37 @@ def run_crew(query: str, model_provider: str, model_name: str, library_type: str
         logger.info(f"🏁 Crew execution finished - delegation cycle complete")
         logger.info(f"📊 Final LLM Stats: {formatting_monitor.get_stats()}")
         
+        # Capture per-agent token metrics for cost attribution
+        logger.info("📊 Capturing per-agent token metrics...")
+        per_agent_metrics = {}
+        agents_list = [
+            ("step_planner", step_planner_agent),
+            ("element_identifier", element_identifier_agent),
+            ("code_assembler", code_assembler_agent),
+            ("code_validator", code_validator_agent),
+        ]
+        
+        for agent_name, agent in agents_list:
+            try:
+                # Get cumulative token usage from the agent's LLM
+                usage = agent.llm.get_token_usage_summary()
+                per_agent_metrics[agent_name] = {
+                    'total_tokens': usage.total_tokens,
+                    'prompt_tokens': usage.prompt_tokens,
+                    'completion_tokens': usage.completion_tokens,
+                    'successful_requests': usage.successful_requests,
+                }
+                logger.info(f"   • {agent_name}: {usage.total_tokens} tokens "
+                           f"(prompt: {usage.prompt_tokens}, completion: {usage.completion_tokens})")
+            except Exception as e:
+                logger.warning(f"⚠️ Could not get token usage for {agent_name}: {e}")
+                per_agent_metrics[agent_name] = {
+                    'total_tokens': 0,
+                    'prompt_tokens': 0,
+                    'completion_tokens': 0,
+                    'successful_requests': 0,
+                }
+        
         # NOTE: Pattern learning is NOT done here!
         # Learning should only happen AFTER test execution succeeds (test_status == "passed")
         # This ensures we only learn from validated, working code.
@@ -261,7 +292,7 @@ def run_crew(query: str, model_provider: str, model_name: str, library_type: str
         if optimization_metrics:
             logger.info("📊 Optimization metrics collected")
         
-        return result, crew, optimization_metrics
+        return result, crew, optimization_metrics, per_agent_metrics
 
     except Exception as e:
         error_msg = str(e)
diff --git a/src/backend/services/workflow_service.py b/src/backend/services/workflow_service.py
index b18b963..81cad76 100644
--- a/src/backend/services/workflow_service.py
+++ b/src/backend/services/workflow_service.py
@@ -78,7 +78,7 @@ def run_agentic_workflow(natural_language_query: str, model_provider: str, model
         
         # Run CrewAI workflow (this takes most of the time - 10-15 seconds)
         # User sees progress messages above while this runs
-        validation_output, crew_with_results, optimization_metrics = run_crew(
+        validation_output, crew_with_results, optimization_metrics, per_agent_metrics = run_crew(
             natural_language_query, model_provider, model_name, library_type=None, workflow_id=workflow_id)
         
         # Stage 3: Generating (50-75%)
@@ -283,6 +283,13 @@ def run_agentic_workflow(natural_language_query: str, model_provider: str, model
                     
                     logging.info(f"📊 Raw CrewAI usage metrics: {usage_metrics_dict}")
                     
+                    # Log per-agent token breakdown
+                    if per_agent_metrics:
+                        logging.info(f"📊 Per-agent token breakdown:")
+                        for agent_name, metrics in per_agent_metrics.items():
+                            logging.info(f"   • {agent_name}: {metrics['total_tokens']} tokens "
+                                       f"(prompt: {metrics['prompt_tokens']}, completion: {metrics['completion_tokens']})")
+                    
                 except Exception as e:
                     logging.warning(f"⚠️ Could not extract CrewAI usage metrics: {e}")
                     # Fallback to empty metrics
@@ -313,6 +320,18 @@ def run_agentic_workflow(natural_language_query: str, model_provider: str, model
                 avg_llm_calls = browser_llm_calls / total_elements if total_elements > 0 else 0
                 avg_cost = browser_cost / total_elements if total_elements > 0 else 0
                 
+                # Prepare token_usage dict from per_agent_metrics
+                token_usage = {
+                    "step_planner": per_agent_metrics.get("step_planner", {}).get("total_tokens", 0),
+                    "element_identifier": per_agent_metrics.get("element_identifier", {}).get("total_tokens", 0),
+                    "code_assembler": per_agent_metrics.get("code_assembler", {}).get("total_tokens", 0),
+                    "code_validator": per_agent_metrics.get("code_validator", {}).get("total_tokens", 0),
+                    "total": sum(
+                        per_agent_metrics.get(agent, {}).get("total_tokens", 0)
+                        for agent in ["step_planner", "element_identifier", "code_assembler", "code_validator"]
+                    )
+                }
+                
                 unified_metrics = WorkflowMetrics(
                     workflow_id=workflow_id,
                     timestamp=datetime.now(),
@@ -345,6 +364,9 @@ def run_agentic_workflow(natural_language_query: str, model_provider: str, model
                     custom_actions_enabled=browser_metrics.get('custom_actions_enabled', False),
                     custom_action_usage_count=browser_metrics.get('custom_action_usage_count', 0),
                     session_id=browser_metrics.get('session_id'),
+                    
+                    # Per-agent token tracking
+                    token_usage=token_usage,
                 )
                 
                 # 4. Record unified metrics