AI Token Burn

Daily token volume across Claude Code and Codex, counted exactly from local logs, plus a labeled estimate band for chat tools that report no tokens. Measured lanes are never summed with estimates.

Daily token burn

Weekly total line: log y-scale; daily cells: log color scale

Less More

What is driving the burn?

Sorted by tool and project; exact lanes plus separated estimate

Categories = tracked project directories (real, from local log paths) — the most defensible honest grouping available. There is no session-content or topic field in these logs, so this deliberately does not infer a work-type/topic taxonomy. Evidence lines below (peak day, active-day count) are real per-project facts, not summaries.

Driver	Trend	Share	Total

Model mix

Share of Claude Code tokens by model — stacked by volume, intensity = tokens/active day

Work phases

Share of Claude Code tool actions — heuristic, not annotated

Mix per week →

Spec verification

Numbers from test/verify output — partial (only saved runs)

Pass / fail per week →

Trend

Exact lanes only — estimates excluded

Recent days

Last 14 active days · exact and estimated columns kept separate, never blended

Scale equivalents

Scale translations, not measured utility/billing/environmental accounting

Rough Fermi (order-of-magnitude) estimates from the current measured total. Each conversion factor is stated plainly in the Basis column — none of this is a billed, metered, or scientifically measured figure.

Measure	Equivalent	Basis