Anacreon

AnacreonUnderstand AnacreonCost and usage

Every token lands on the work it bought.

Anacreon doesn't ask agents what they spent. It reads usage from each AI tool's own logs, counts it against whatever the session was working on, and uses finished work to size the work still to come.

Level 3 · Cost and usage

Read, attribute, learn. Usage is read from the tool's logs. It is counted against every claim and request the session held. Finished work calibrates the estimates for new work of the same size.

From a log line to an estimate

  1. The tool's own logClaude Code's transcript, Codex CLI's session log, or Anacreon's runner.
  2. A usage reportTokens by model: input, cache write, cache read, output. Labelled with where it came from.
  3. AttributedTo every piece, chunk and request the session held, and to the member and model.
  4. ActualsEach finished item ends with what it really cost, in tokens and in hours.
  5. CalibrationFinished items of one size become the ranges new estimates of that size use.
  6. What's leftThe project's remaining work, rolled up, with the path that can't run in parallel.

How a session's tokens are counted

A session reports its usage as it goes. Each report is counted toward everything the session held when the report arrived.

Usage counted against what a session held One session over time. It holds a claim on piece A, then also takes a review B, then finishes A and keeps B. Four usage reports arrive: the first counts toward A, the second and third toward both A and B, the fourth toward B only. One session, over time Piece A claimed Review B claimed Usage reports A A + B A + B B
Each report counts toward every claim held at the time. When calibrating estimates, a report shared by two claims is split evenly between them. Usage while a session holds nothing is shown on its own line, so nothing disappears.

Where each number came from

Every usage row says how Anacreon got it. Only usage read from a tool's own logs counts as evidence.

SourceHowCounts toward actuals, estimates and prices
Read from the tool's logsA small adapter on the member's machine reads the AI tool's own log and posts what it recordedYes, verified
Self-reportedA tool whose logs can't be read states its usage, with a one-line basisNo. Shown, labelled, kept apart
  • A session that reported a model it didn't declare, a model from a vendor the project doesn't allow, or one the work's requirement doesn't accept, is flagged on the Record page with the reason.
  • An hour after a session starts with no verified report, the session is told its usage hooks may be missing and given the command to install them.

Estimates that learn

An estimate is sized by what the work touches, not by a guess at hours. The size picks a class, and the class gives a range of tokens and of time.

ClassStarting definition
XSOne or two files in one area. No migration, no open decision, no unknown.
SA handful of files in one area, with a pattern to copy. At most one small unknown.
MUp to about ten files in one area, possibly one migration. At most one open decision or unknown.
LMany files across two or three areas, with migrations, and a couple of open decisions or unknowns.
XLCrosses several areas, with several migrations, open decisions or unknowns. A candidate to split.
Estimate ranges by size class, with actuals An illustration with no real figures. Each size class has a typical range and a longer worst-case range, growing from XS to XL. Dots show finished items' actual cost, which is what the ranges are calibrated from. Tokens per item (illustration, not real figures) XS S M L XL more tokens →
Solid: typical range. Outline: worst case. Red dots: finished items' actuals. Until a class has enough finished work behind it, its estimates are marked uncalibrated and use the owner's starting ranges.
  • Every estimate carries a confidence and what would move it.
  • A new estimate is a revision that points at the one before. Nothing is overwritten.
  • Estimates only advise. Nothing is hidden or refused because of one.
  • Estimates live only inside Anacreon. They are never copied into commits, pull requests or your repository's documents.

Dollars, on one scale

Usage is stored in tokens by model. Dollars are worked out when a page is shown, from the project's price table of each model's list price, and never stored. A subscription plan and an API key land on the same scale, so the cost of the work compares however each person pays.

The roll-up adds up what is left from each open item's latest estimate, in tokens by model and in wall-clock time, and finds the critical path: the part of the work that can't run in parallel.

Coming next for cost

WhatWhy it mattersStatus
Richer usage reports: subagents, cache lifetimes, effort, timing, compactionsShows where tokens go inside a session, not only how manyIn design
Every tool call recorded: tool, sizes, time and outcomeShows which parts of the workflow cost the most contextIn design
One export of usage, calls and outcomesAnalyse a project's AI spend in your own toolsIn design
The exact rules: commitments and usage Deeper
  • Usage is tokens by model in four kinds: input, cache write, cache read and output. Time is wall-clock from claim to done.
  • A report is keyed on the session and the tool's own message id, so a retried report never counts twice.
  • A token commitment counts input, cache-write and output tokens, not cache reads. Estimates and calibration count every kind.
  • A member may share their AI plan when they join. Nobody has to.
  • Tokens aren't the only cost. People's reviews and answers are often the real bottleneck, which is why questions arrive ranked by how much work they hold back.
  • A transcript reported to two projects counts in both.