usage-data

Public, auto-refreshed data file powering maxghenis.com/usage.

usage.json is regenerated every ~30 minutes by build_usage.py + extract.py, which read the raw Claude Code and Codex session JSONL on my machine and aggregate to daily totals. Only daily aggregates are published: token counts, computed cost at public API list prices, and model breakdowns. No message content, no file paths, no project names.

Methodology

• Dated to when usage happened. Every API call is bucketed by its own timestamp (UTC), not by session start, so sessions that span weeks don't pile onto their start date.
• Bill-faithful codex accounting. Codex fork/resume writes a snapshot file containing a full replay of prior history under a fresh session id. Replays are detected structurally — the leaf session_meta (carrying forked_from_id) adjacent to the copied ancestor's session_meta, with the native boundary at the first task_started whose preserved started_at agrees with its own outer timestamp — never by wall-clock timing, which misses multi-second copied prefixes and falsely discards fresh same-second work. Genuine cumulative-counter resets segment a file into billing epochs that are summed; rate-limit-only heartbeats are ignored. Only live continuation deltas of total_token_usage count.
• Claude requests deduplicated globally by (message.id, requestId), keeping the most complete record (largest token total): subagent transcripts stream several usage snapshots per request (the first is a placeholder with output_tokens≈1; the last carries the billed totals), resumed sessions copy history into new transcript files (usually verbatim, occasionally truncated), and the pick must not depend on which file sorts last.
• Cache-creation captured. Claude prompt-cache writes are priced at the 5m/1h write rates; cache reads at the cache-hit rate.
• Pricing is pinned per model generation at public list prices (standard tier, no Batch/Flex modifiers; Claude 4.6+ have no long-context premium), from Anthropic's pricing page and, for OpenAI, cross-checked against LiteLLM's pricing table. A price row applies only to its own model id (plus date stamps and Codex variants such as -codex), so a new generation such as claude-opus-5-5 can never inherit an older row's rates; a model with no row is published at $0 and named under pricing.unpriced in usage.json. Claude fast-mode requests are priced at their own fast-mode rows.
• Durability. A local scan cache (keyed by size, mtime, and accounting- algorithm version) retains per-file results after Claude Code rotates transcripts (~30 days) or codex archives sessions. For days whose transcripts rotated before the cache existed, claude_history_seed.json carries the Logpile ledger's deduplicated daily totals; per day, whichever source knows more tokens wins wholesale.
• Coverage: ~/.codex{,-2,-3,-4}/{sessions,archived_sessions}, OpenClaw codex homes, and ~/.claude/projects (including subagents/ transcripts).

Claude figures match ccusage 20.0.24 run with --timezone UTC to the cent for July 2026, and day by day for September 1–27 (see RECONCILIATION.md, 2026-09-28).

This methodology reconciles month-by-month with Logpile's adversarially reviewed strict accounting (session_daily_effective, native columns) — exact on most months, within ~2% elsewhere, with the residual fully attributed to private sessions the ledger intentionally excludes. It intentionally diverges from ccusage-class counters on Codex, which both count inherited fork history whose replay spans multiple wall-clock seconds and discard usage after genuine counter resets; see RECONCILIATION.md.

Human/automated split comes from Logpile session-origin classification.

Tests

uv run --no-project --with pytest --with hypothesis python -m pytest -q

tests/test_accounting_properties.py holds the property tests: every billed Claude request is counted exactly once at its billed vector, per-day totals equal the sum over unique requests, the result is independent of file order, and no model id is priced at another generation's rates.