aiUsage CLI turns Claude Code and Codex logs into CSV

The tool exists because its author writes a monthly article on AI usage and, each time, had to dig through saved logs, total tokens per model, and convert them at API rates. aiUsage compresses that routine into one command, and the distributed build needs no Rust, Python or API key.

Its design choices follow from that reporting purpose. Output is one row per model with no total row, on the assumption that totals, graphs and comparisons against subscription fees belong in Excel or Google Sheets. Amounts come from a pricing table bundled into the executable, which is why they are labelled a reference conversion rather than an actual subscription payment, and why models missing from that table get tokens but a blank amount plus a terminal warning.

Accuracy work sits mostly in the cache handling. Claude Code logs distinguish five-minute and one-hour cache writes and aiUsage reflects that; Codex treats what is left after subtracting cache reads from input as normal input, and its reasoning tokens are already inside output so they are not added again. Repeat updates to the same Claude Code response and Codex's cumulative counters are not simply summed, with Codex differences anchored on records from before the period start. That last point is probably where the numbers are most sensitive, so users comparing months may want to check the bundled pricing table and how a period boundary falls before treating the figures as final.