What you get
- Every card carries a cost pill — price, token sum and the Claude model family — so you see which tasks burn budget without opening them.
- A task that used several models shows the dominant one with a
+; hover for the full per-model breakdown. - Epic parents add up their own and all their subtasks' costs, so the parent's pill is the Epic total.
- The detail panel shows a Tokens & Cost block with total cost, total tokens and a per-model split.
How you use it
- Nothing to set up — cost tracking is on automatically per project.
- Work tasks as usual; usage adds up per task across replies and sessions.
- Open a task's detail panel to see the per-model breakdown.
How the cost is calculated
The Tokens & Cost block breaks each model's usage into four token buckets. Every reply Claude sends reports how many tokens it read and wrote, and Sightdeck sums them per model:
- In — input tokens: your prompt, the task context and files sent to the model.
- Out — output tokens: everything the model writes back.
- Cache W — cache write: the slice of the prompt the model stores so it doesn't have to re-process it next turn. Written once, billed a bit above the input rate.
- Cache R — cache read: the cached slice read back on every following turn instead of being sent fresh. Billed at roughly a tenth of the input rate.
The cost is the sum of all four buckets, each multiplied by that model's list rate per million tokens. All four count toward the price — including cache reads. So the total you see is the true theoretical value of the work at list rates.
The Tokens column is different: it counts only in + out + cache write and leaves cache reads out. That's deliberate. The cached prompt prefix is re-read on every turn of a long session, so adding those reads up balloons the count into the millions and buries how much real work a task actually did. Cache reads stay visible in their own Cache R column — they just don't inflate the headline token number.
Two edge cases: a model whose list price isn't published yet shows $– instead of a fake $0.00, and costs recorded after a task was already finished are flagged as estimated. The rates are indicative Anthropic list prices, not your actual invoice, and can change over time.
Pairs well with model selection
Cost tracking shows you where the budget goes; model selection keeps it in check by matching the model to each task's complexity. Watching both together is how you keep heavy parallel runs affordable.
