Run opencues usage. It prices every call your running hosts have made, per model, with token counts and cache-hit rates, and gives you a total.
| Command | What you get |
|---|---|
opencues usage | calls, tokens, cache-hit rate and cost per model, plus a total |
opencues usage --json | the same, machine-readable |
opencues usage --reset | clears the counters and starts again |
Why a per-feature estimate is not enough
Several features can fire on a single edit. Word cues, sentence cues, the two session cues and any blank you trigger each make their own call, so reasoning about one feature at a time understates what you are actually spending. This counts everything, which is the number that matters.
Cache discounts are included
Most of what OpenCues sends is a long, stable instruction block that providers can cache, and a cached token is far cheaper than a fresh one. The cost shown reflects that split rather than pricing every token at full rate, so the figure matches your bill rather than a worst case.
It costs nothing to look
The command makes no LLM calls. It reads snapshots the hosts have already written, prices them, and prints. It also counts the short-lived background reader behind session-contradiction cues, which runs outside the editor and would otherwise be invisible.
One gap
Chrome is not counted. The extension has nowhere to write a usage snapshot, so calls made in the browser do not appear in the total. Everything running in a terminal does.
Spending less
Every feature that costs anything can be turned off in ~/.cues/OPENCUES.md, and the passive cues are the ones worth reviewing first if the total surprises you. You can also point a bucket at a cheaper model: cues, auditors and blanks are routed separately, so you can keep a fast model for the ones that run constantly and a stronger one for the ones you invoke by hand.
Related: Best model for inline AI · Contradict an earlier decision · Ask-cues · GitHub repo