Where Am I Burning Tokens?v1.1.0

At the end of each month, an AI engineering team reviews token usage and costs across categories. They discover paperclip logging consumes 36% of tokens but only 5.8% of costs, indicating a cheap model used heavily for context logging, which could be optimized or disabled to free up budget.

During a weekly check, an admin sees cron services account for 48% of costs but only 25% of tokens. This flags an expensive model being used for heartbeats or scheduled tasks, prompting a switch to a cheaper model like Haiku or a local model to reduce costs by over $20 per week.

A company runs many sub-agents for parallel task processing. The token audit reveals sub-agents use 28.8% of tokens and 9.4% of costs, which is efficient. However, if a sub-agent category showed high cost with low tokens, it would indicate an expensive model choice that needs adjustment.

An AI agent uses Paperclip for logging all interactions. The analysis shows Paperclip consumes over a third of tokens, mostly for reading large contexts with little output. The team decides to reduce logging detail or disable Paperclip entirely, saving nearly $6 per week in token costs.

Before scaling AI agent usage to more users, the operations team runs a token audit to understand current consumption patterns. They see main interactive sessions use only 10% of tokens but 36.5% of cost due to Sonnet model, helping them budget for higher volume of expensive interactions.

Offer monthly subscriptions with token limits. Use the token audit to monitor per-category consumption and alert users when they approach limits. Provide upsell options for higher caps or cheaper model tiers for background tasks like cron and logging.

Charge customers based on actual token usage or per session. The audit provides transparency into which sessions are driving costs, enabling fair billing and helping customers optimize their own usage. Offer discounts for using cheaper models for non-critical tasks.

Provide consulting services to enterprises to analyze their AI agent token consumption using this skill. Identify inefficiencies like sinkholes or expensive cron jobs, and recommend model changes or configuration adjustments. Charge a fixed fee per audit or a percentage of cost savings.