Skip to main content
Settings
Color Mode
Theme Skin
Background

Appearance preferences are saved in this browser only.

Environment
Current Environment Production

Built with JEKYLL_ENV=production. Changes require deployment.

Theme & Build
Jekyll v3.10.0
Last Build Jul 15, 04:28
Page Location
Page Info
Layout default
Collection docs
Path _docs/ai-usage.md
URL /docs/ai-usage/
Date 2026-07-15

AI Usage & Cost

Every token the robots spend building lifehacker.dev — metered per call, attributed per PR, published here. Radical transparency, mildly embarrassing totals.

Table of Contents

AI Usage & Cost

This site is written, reviewed, fixed, and triaged by AI agents. That effort is not free — it’s measured in tokens — so we meter it: every model call in the pipeline records what it consumed, every pull request carries a running cost comment, and the nightly ledger folds it all into the numbers below. Nothing here is hand-typed.

How to read the dollars: every figure is API-equivalent — what the tokens would bill at Anthropic’s list prices. The pipeline authenticates with Claude Code subscription auth (OAuth) first, so the marginal dollar cost of those runs is zero; the API-equivalent figure is the honest way to compare effort anyway. If a run ever falls back to a metered API key, that share is real spend and is broken out below.

Last rollup: 2026-07-15T02:55:35Z · 1 AI calls on record

Totals

windowcallsoutput tokenscache read/writecost (API-equiv)
all time18141305525 / 28245$0.6397
last 30 days18141305525 / 28245$0.6397
last 7 days18141305525 / 28245$0.6397

Who spends it (last 30 days)

workflowcallsoutput tokenscost
pipeline18141$0.6397
By role and by model
rolecallscost
content-reviewer1$0.6397
modelcallsoutput tokenscost
claude-opus-4-818141$0.6397

What a pull request costs

A PR’s bill has two halves: the run that wrote it (creation), and everything the machine did to it afterward — the harness, reviews, auto-fixes, brand adjudication (downstream). Each PR carries its own live version of this table in a sticky comment; these are the all-time heavyweights.

PRcallscreationreviews & fixestotal
#3101$0$0.6397$0.6397

Spend by month

monthcallsoutput tokenscost
2026-0718141$0.6397

Auth mix

  • oauth: 1 calls, $0.6397 API-equivalent — subscription-covered, $0 marginal

How the meter works

Every model call in this repo flows through one runner (scripts/ai/run.sh), which asks Claude Code for a JSON result and records the usage payload — tokens in, tokens out, cache traffic, reported cost — as one JSONL record. The end of each AI job publishes those records three ways: a step summary on the run, an ai-usage-* artifact, and a sticky cost comment on the PR it worked on. A nightly sweep folds the artifacts into _data/ai_usage/ledger.jsonl and regenerates this page’s data. The full design doc lives in docs/AI-USAGE.md.

The meter can’t see two things, and says so: a run that crashes before writing its result payload leaves no record (the record notes failures that finish), and work done on a laptop outside CI stays off the books. Both are documented gaps, not surprises.