Pi is already the lean option, with four core tools and minimal context overhead. Its remaining cost levers are model choice (switch mid-session to cheap models for routine stretches) and session hygiene via branching and compaction.
Pi's 30+ provider support with mid-session switching makes 'cheap for routine, frontier for hard' a live command instead of a config project.
Session branching lets you explore without dragging the full history forward; compaction keeps long-lived sessions from becoming token barges.
Pi's core is lean; extension sprawl is where hidden calls creep in. Know what each installed extension sends and when.
The provider-agnostic tactics (prompt caching, retry budgets, batch APIs) are in the general playbook.
Per request, usually. Minimal tool definitions and lean context mean less fixed overhead. Total cost still tracks how you use it: model choice and session length dominate, same as anywhere.
A value model as default with mid-session escalation for hard moments, or a flat-rate endpoint in models.json if Pi is your daily driver, which removes the question entirely.
The structural version of all of this: run Pi on a flat monthly price, and the bill stops being a variable to manage. 2-minute Pi setup → · Best models for Pi →