tokens

How this is counted

Claude Code writes a JSONL transcript for every session under ~/.claude/projects and keeps it forever. Codex writes rollout files under ~/.codex/sessions but prunes old ones. A script that simply rescans the disk would therefore watch its own lifetime total shrink over time, which is a strange property for a lifetime total.

So the scanner keeps a ledger outside the repo, keyed by session, day and model, where a bucket can only ever be raised, never lowered. Once a day is recorded it survives the deletion of the log that produced it. On top of that sits a manual baseline taken from the Codex profile page, which is server-side and knows about machines and sessions that vanished from disk long ago. Only Codex activity dated after the baseline is added to it, so nothing is counted twice. Claude Code carries two retired-machine baselines the scanner can no longer reach: one measured before that laptop left the rotation, one an honest estimate from a machine whose logs are gone — marked as such, because a lifetime number that quietly forgets a whole laptop is worse than one with a labelled estimate in it.

The heatmap below is narrower than the headline on purpose: it can only show days the ledger has actually observed, while the headline includes pre-ledger history recovered from the baseline. Anthropic publishes no lifetime counter at all, so the Claude Code and Cowork figures are local logs and nothing else. The Codex scanner has been cross-checked against Codex's own local database and agrees to the exact token. Totals include cache reads, which dominate. Claude and OpenAI also tokenize and report their own model traffic, so the vendor totals are native accounting units rather than a strict cost comparison. This page measures raw throughput, not a bill or subscription-limit meter.

Scope: this counts coding agents — Codex, Claude Code, and Cowork. Plain web chat on chatgpt.com and claude.ai is excluded for a boring reason: neither exposes a token count anywhere, so any number for it would be made up. If either ever publishes one, it gets a baseline entry here.