Where did my tokens go, per session and per model?

Usage by model

Muster reads token counts from the CLI’s transcript, so the numbers are the CLI’s own accounting, per session and per model.

Cost per turn

The panel plots the cost of every assistant turn in the selected session, and the ratio between the last quarter of the session and the first — how much more the same request costs now than it did at the start. Below eight turns it shows only the curve; the ratio stays quiet, because a ratio from four points says nothing.

It does not predict the next turn, and it does not estimate what restarting would cost. Nobody has measured either.

Cache misses

An amber band on the curve marks turns where the cache expired and the whole conversation was rewritten instead of read — the same tokens at 1.25× instead of 0.1×.

This is a derived signature, not a field in the transcript. On a subscription the cache lives about an hour, so coming back from a long break with the session still open lands here. If the band shows up repeatedly, the fix is shorter sessions, not faster typing.

Per model

A session records which model produced each turn. A role can pin a model, so “the cheap model does the mechanical tasks and the expensive one does the design work” is a thing you configure once on the role rather than remembering per session.

The efficiency panel

Under the queue: four bands of time — model working, tool running, waiting for you, idle — plus context reuse and recommendations tied to a measured number rather than a rule of thumb. It is the view that answers “where did the last seven days go”, and it is per project.