Skip to main content

​
Spend

Cost by day, by model, by project — and cost filtered by a search query, which is the one no dashboard can do. Prompt-caching gets its own card, because cache reads and cache writes are priced very differently from fresh input and are usually where the money actually goes.

​
Always an estimate

Turnlog computes cost from the token counts in your own logs against a shipped price table covering Anthropic and OpenAI rates. It is labelled an estimate everywhere, because it is one: it is not your bill. Cursor IDE sessions are the exception — they use the cost Cursor itself recorded. Rates wrong for you — Bedrock, Vertex, an enterprise agreement, or plain disagreement? Override any model with modelPricing.

​
Two ways naive counting gets this wrong

Both are worth knowing, because they are why a hand-rolled script disagrees with Turnlog. Usage repeats per line, not per response. Claude Code writes one JSONL line per content block, and every one of those lines repeats the same message.id and the same usage object. Summing per line inflates cost by roughly 2.7×. Turnlog counts usage once per response. Resuming copies the whole history forward. A resumed session’s file contains everything from the previous one, so counting per file bills the same conversation repeatedly. Turnlog dedupes across a resume chain: the first occurrence of a message wins.

​
Calendar

Your sessions placed in time — a week timeline with sessions as blocks at their real hours, or a month heat-map. Colour by project or by agent; whichever you choose fills the block and the other becomes its edge stripe, so both encodings are always present.