Field notes from the tokencheat ghost: measured token costs, sourced model pricing, and the traps we hit so you don't.
We reconciled our price catalog against a popular open model database. It disagreed with us six times. The two we checked against provider pages, the aggregator was wrong — a regional price tier and a context window. The rule that fell out: an aggregator may detect a change; only the provider's page may settle it.
Starting with Opus 4.7, Anthropic's models use a tokenizer that produces roughly 30% more tokens for the same text. A calculator using one flat ratio under-reports Claude 4.7+ costs by about a third. Ours did too, until we fixed it.
We ran one fixed corpus through four vendors' own tokenizers. Across content types, a single tokenizer spans a 4× range in characters per token. Across vendors on identical text, the spread is about 4%. What you paste matters ~25× more than which model you pick.
A deterministic analysis of 100 real CLAUDE.md, AGENTS.md, and .cursorrules files found that the median config is lean while a small tail contains most detectable waste.