LLM Spend Receipt

The problem

Usage dashboards are often buried behind auth walls or cluttered with irrelevant metrics. Developers lack a simple, automated way to see exactly where their daily LLM budget went without manual reconciliation.

What it does

• Parses provider usage CSVs or API responses into a phone-sized text receipt.
• Calculates real cache-hit rates to show actual savings versus list prices.
• Identifies "blow-up" events where a specific model or tag consumed a disproportionate share of the daily budget.
• Generates actionable caps, such as hard-cap suggestions or prompt reordering for better caching.
• Distinguishes between cache-miss, cache-read, and cache-write token buckets for accurate pricing.

Frameworks & tools

Python 3.10+ standard library (csv, pathlib). Supports any provider exporting CSV usage data (Anthropic, OpenAI, etc.).

Why this beats prompting it yourself

General-purpose prompts often fail at token bucket math, incorrectly pricing cached tokens as standard input tokens. This skill enforces strict normalization across different header aliases and ensures $0 rows from local or seat-based models are retained for an accurate view of total system load.

Use cases

• Automate a morning Slack/email notification showing yesterday's total model spend.
• Audit a large agent run to see if caching strategies are actually reducing costs.
• Identify runaway jobs in a multi-tenant environment using tag-based grouping.
• Set baseline 'typical spend' alerts to catch billing spikes before the invoice arrives.

Known limitations

It is a daily operator view, not an accounting tool for invoice disputes or tax reconciliation. It does not perform multi-month forecasting.