Tokens and the reasoning budget
Tokens are the unit AI usage is billed in. The reasoning budget sets how long it thinks at each step.
An analogy
Like hiring a contractor. A quick look and start is cheap. Measuring, drawing plans and a second check costs more, but you redo less.
In pstack
pstack sends out helper AIs and has several models review together, and all of that costs tokens. So by default it's slower and pricier, but more careful. /pstack:setup-pstack asks you to pick a tier. large is the default (xhigh). unlimited runs every model at its highest. medium and small cost less.
Three ways to save. Use a cheap, fast model for the coding roles. Shrink the review panel. Skip /pstack:poteto-mode for small changes.
Common misconceptions
- "
autois a model name": it means "no specific model; use the same one as the main chat".inherit-parentmeans the same. - "A higher budget is always better": hard problems deserve a high budget. For everyday changes it's just slower and pricier.
Try it: Keep costs down
Related: Model roles and the pstack-models config, Subagents