Investigate high Bedrock token usage expenses

[mxiamxia]

Hi @awsapm, can you take a look the GenAI tokens usage in my services? why the cost is high?

[github-actions]

🎯 Application observability for AWS Investigation Complete

Investigation completed successfully! Here are the results:

---

High Token Usage Analysis

Primary Cost Drivers:
• Pet Clinic Agent: 8.7M output tokens, 58.4M input tokens (39,125 invocations)
• Nutrition Agent: 5.6M output tokens, 25.4M input tokens (21,671 invocations)
• Customers Service: 409K output tokens, 15K input tokens (1,703 invocations)

Cost Breakdown by Model:
• Claude 3.5 Haiku: Extremely high usage with ~1,500 avg input tokens and ~225-270 avg output tokens per call
• Titan Text Express: Lower per-call usage but still significant volume

Key Issues Driving High Costs:

1. Excessive Input Token Usage: Average 1,500+ input tokens per Claude call suggests large context windows or inefficient prompting
2. High Invocation Frequency: 60,796 total GenAI calls in 24 hours (~42 calls/minute continuously)
3. Agent Orchestration Overhead: Multiple agents calling each other creates token multiplication

Immediate Cost Reduction Recommendations:

• Optimize Prompts: Reduce average input tokens from 1,500 to <500 per call
• Implement Caching: Cache frequent responses to avoid repeated expensive calls
• Batch Processing: Group similar requests to reduce individual API calls
• Model Selection: Consider switching to cheaper models for simple tasks
• Rate Limiting: Implement throttling to control invocation frequency

The agents appear to be running continuously with very high token consumption per interaction, suggesting optimization opportunities in prompt engineering and request batching.

---

✅ Status: Complete
👤 Requested by: @mxiamxia
🔗 Workflow: View details

Powered by AI Agent