How to Check Pi Agent Usage and Cost

Pi — the pi-mono agent by Mario Zechner (badlogic) — takes a deliberately spartan view of the terminal: a tight loop, a small surface, and almost no dashboard. That minimalism is the point, but it leaves a blind spot. Pi will happily chew through a long editing session without ever telling you how many tokens went out the door, and because it is bring-your-own-key, the bill lands on your provider account days later with no per-session story attached. The good news: Pi does write a usage trail to disk, and that trail is enough to reconstruct the whole picture.

How do I see my Pi agent usage?

You do not need to bolt anything onto Pi or change how you run it. One command reads what is already on disk and prints a breakdown:

$ npx whoburnedmore↳ probing sources… claude codex gemini pi …↳ pi: parsed 312 log entries across 11 sessions   PI AGENT USAGE  ────────────────────────────────────────  2026-06-17  input  1,204,800  output   96,300  2026-06-16  input    742,100  output   58,900  2026-06-15  input  1,933,400  output  151,200   BY MODEL (your keys)  your-primary-model      6.8M tokens  your-fast-model         2.1M tokens   7-day total:  9.0M tokens   est. at your rate

whoburnedmore lists Pi among the terminal-native agents it knows how to read. It walks Pi's local log entries, pulls the input and output token counts off each turn, and folds them into daily, per-model, and per-session views. Nothing about the session — no prompts, no file names, no code — is needed for that math, so none of it is touched.

Why Pi hides the number in the first place

Heavier agents ship a status bar, a budget meter, or a /cost command. Pi deliberately skips that furniture to stay fast and small. The trade is that spend visibility becomes your job — which is exactly the gap whoburnedmore fills by reading the logs Pi keeps regardless.

What does the Pi agent cost?

Pi itself is free and open source — there is no Pi subscription. What you actually pay for is the model behind it, billed straight to whatever provider key you handed Pi. So “what does Pi cost” really means “what did the model I pointed Pi at charge me,” and that depends entirely on which model and which provider you chose. Because the rate is yours, we will not invent a per-token figure here; instead, whoburnedmore applies the rate that matches the model recorded in your own logs.

Two consequences fall out of that. First, swapping Pi onto a cheaper model changes your cost without changing a single thing about how you use the agent. Second, because output is usually the pricier half of any provider's rate card, a Pi habit that produces a lot of generated code costs more per token than one that mostly reads and answers — and the per-model split makes that visible.

Where Pi keeps its usage trail

1. 1

   Local log entries, written per turn

   Pi records each model exchange to its own log files on your machine. Those entries carry the token counts whoburnedmore needs — no telemetry endpoint, no cloud lookup.
2. 2

   Parsed in place, never moved

   The reader opens those files read-only, extracts only the numeric fields, and leaves the logs untouched. You can run it as often as you like without side effects.
3. 3

   Aggregated into day / model / session

   From those raw turns whoburnedmore builds the daily totals, the per-model split, and a per-session view for tracing a single pricey run.

How do I compare Pi to my other agents?

Few people run only Pi. You might reach for it on small, fast jobs and a heavier agent for big refactors — which makes the real question not “what did Pi cost” but “how does Pi's share stack up against the rest of my week.” The same command that reads Pi reads the others, so they all land in one table and one chart 📊:

Seen on its own, Pi's number is just a count. Seen as a slice, it answers something more useful: is Pi actually your lean option, or has it quietly become a chunk of your spend because you pointed it at an expensive model? Lining the agents up side by side is the only way that question gets a real answer.

Does reading Pi's logs touch my provider key?

No. Your provider key lives in Pi's own configuration and is never read, needed, or transmitted by whoburnedmore. The reader only wants the numeric usage fields — token counts, model identifiers, and timestamps. Run npx whoburnedmore --dry-run to print exactly what a submission would contain, or npx whoburnedmore --local to keep the whole breakdown on your machine with nothing leaving at all. 🛡️