Fixes reports that read as inflated while undercounting cost. ccusage-fleet's own aggregation was already correct — verified against raw transcripts on a four-host fleet — so this release targets an upstream over-count, pricing defaults, and presentation.
ccusage 20.0.19 corrects a Codex over-count
The bundled ccusage moves from 20.0.17 to 20.0.19, which fixes codex: dedupe replayed usage, stop double-counting reasoning in derived token totals, and correct replay and pricing accounting.
Expect reported totals to fall after upgrading. On one Codex-heavy host over an identical date range, the same transcripts went from 24.67B tokens / $18,299 under 20.0.17 to 14.56B / $12,273 under 20.0.19 — 41% fewer tokens. Nothing was lost; 20.0.17 counted replayed Codex turns more than once. Fleet-wide the drop was 122.2B to 97.4B.
Pin the old behaviour with --ccusage-version 20.0.17 if you need to reproduce earlier reports.
Online pricing by default
ccusage defaults to online pricing; ccusage-fleet had flipped it to offline, so any model newer than the snapshot bundled with ccusage was silently priced at $0. That dropped 9.7% of tokens (11.9B, mostly claude-opus-5) out of the cost column on the fleet under test; one day priced $0 offline and $32.14 online. After this release, unpriced tokens on that fleet fall to 0.4%.
--offline still selects the bundled snapshot. If you rely on it, expect $0 for recently released models.
--graph-metric output
Cache reads were 96.4% of total tokens, so token-scaled bars track context replay rather than work and diverge from cost. Scale bars by output tokens instead — the closest proxy for work performed.
Explanatory notes under the table
Note: lines now report the cache-read share, models missing from the pricing table, and the reasoning tokens ccusage folds into totalTokens without a column of their own. --json gains a notes array; schemaVersion is unchanged.
Full changelog: v0.3.0...v0.4.0