LLM Bill Analyzer

Drop the usage CSV from your Anthropic or OpenAI console and get the analysis the invoice never gives you: spend by model and day, your actual cache hit rate, what prompt caching would have saved, and what routing part of the traffic to a cheaper model is worth.

# Your file never leaves this tab. No upload, no server, no analytics on the contents. The parsing is ~300 lines of vanilla JS on this page. View source.

Drop your usage .csv here or click to choose a file
Anthropic Console → Usage → Export · OpenAI Platform → Usage → Export
The demo is generated, clearly fake traffic, so you can see what the analysis looks like before trusting it with real numbers.
Column mapping

Auto-detected from your headers. Fix anything that guessed wrong, then re-run. Export formats drift; the mapper is the warranty.

Totals

By model

By day

Caching: claimed vs left on the table

Routing: the cheaper-model question

Pick a share of your most expensive model's traffic that a smaller model could serve (summaries, classification, short answers), and see what the move is worth. Whether quality holds is an evals question; this is the money side.

Assumptions, stated plainly: estimated costs use the hand-verified price table from the LLM Model Reference (standard tier, verified ); models the matcher does not recognize are totaled from your cost column when present, otherwise flagged. Cache savings assume Anthropic-style economics (cache reads at 10% of input price; the one-time 25% write premium is ignored, which slightly flatters high-churn caches). Daily-aggregate exports cannot show per-request shapes, so the routing estimate reprices tokens, not requests. This is a planning tool, not an invoice. The engine is open source, tested, with a CLI: github.com/AlexRyan92/llm-bill-analyzer.