By name: usage-cost description: > Estimates what this machine's Claude Code usage would cost on each plan -- pay-as-you-go API pricing, Pro, Max 5x, and Max 20x -- by replaying the exact per-request token usage in local transcripts (~/.claude/projects) through each plan's rolling 5-hour and 7-day rate limits, then printing a table of flat fee plus estimated overage as a monthly total, ordered by plan size. Use when the user asks "which Claude plan is cheapest for me", "is Max worth it", "am I on the right plan", "what would my Claude Code usage cost", or "subscription vs pay-as-you-go". Do NOT use for live Anthropic Console billing, claude.ai web/desktop/mobile usage, other machines, or direct API-key logs -- it only sees this machine's Claude Code transcripts. allowed-tools: Bash(uv run *)
Claude Code Plan Cost Comparison
Claude Code transcripts record the exact billed usage object from every API response
(input_tokens, output_tokens, cache_read_input_tokens, cache_creation_input_tokens
with 5m/1h TTL split). The script replays those requests as timestamped, API-priced events
and runs each plan's rolling rate limits over them:
monthly cost = flat fee + overage (usage beyond the 5h/7d limits, billed at API rates)
It prints a table ordered by flat fee — Max 20x, Max 5x, Pro, pay-as-you-go — with no
overage cap modelled (caps only block you; raising them is the norm), so total is the true
cost of letting all your usage through.
Usage
uv run --script SKILL_DIR/scripts/usage_cost.py
The script declares its one dependency (Click) inline, so uv run --script installs it
automatically — nothing to set up. Options:
--days N— only the trailing N days (e.g.--days 7for this week's pace)--breakdown— also print per-model and per-day cost--projects-dir PATH— scan a different transcript root--pro-5h $ / --pro-7d $— override the calibrated Pro limits (Max tiers scale 5x/20x)
How the plan limits are calibrated (read before trusting the numbers)
Anthropic publishes plan limits only as vague "prompts per 5 hours," useless for a dollar
model. The script instead carries limits expressed as API-$ of usage per rolling window,
calibrated empirically (constants PRO_5H, PRO_7D at the top of the script):
- Pro 5h limit (
$20) was solved so the simulation reproduces a real Pro overage bill — the robust anchor, since it integrates many days of bursts. The fit came out ≈$20.6; it's rounded to a round$20(which also happens to equal the $20/mo Pro price — a coincidence, not a mechanism). Treat the precision as illusory. - Pro 7d limit (
$875) came from a live/statusweekly percentage (softer; the 7-day window rarely binds — for most users overage is driven entirely by 5-hour bursts). - Max 5x / 20x scale Pro's limits by 5x / 20x, matching Anthropic's own framing.
These approximate Anthropic's opaque internal limits and assume an Opus-heavy, cache-heavy
mix. The absolute overage figures are soft (call it ±50%); the ranking is what to trust.
Two independent code reviews (Codex + Opus) pinned the only fragile boundary: Pro never
wins, and Max 5× stays cheapest unless the true 5h limit is ~⅓ below this estimate (then
Max 20× wins) — and both calibration anchors ($20.6 bill, $33 /status) sit at or above $20,
so the estimate is more likely low than high, which only favors Max 5×. For a very different
usage mix, recalibrate via --pro-5h / --pro-7d (from your own overage bill or /status).
How to read it
- Burst shape, not raw total, drives cost. $800 spread thin can fall entirely under a cheap plan; $800 in one 5-hour window blows any plan's session limit. That's the whole point of the rolling simulation — it captures this where a monthly sum can't.
- Pay-as-you-go is the raw API value of everything; subscriptions are typically a 5–10x discount, so the real question is which subscription.
- The cheapest plan is flagged. The sweet spot is usually the smallest plan whose limits absorb your bursts with near-zero overage.
Caveats (also printed by the script)
- This machine's Claude Code only — no claude.ai web/desktop, other machines, or Console billing. Calibration assumes that's the bulk of your usage.
- Transcripts are pruned after ~30 days, so older history is a floor, not a total.
- Records predating the cache TTL split are priced at the cheaper 5m rate.
Pricing freshness
Rates are hardcoded in BASE at the top of the script (as of 2026-06: Opus $5/$25,
Sonnet $3/$15, Haiku $1/$5, Fable $10/$50 per MTok; cache read 0.1x, write 1.25x/5m & 2x/1h;
no >200K long-context premium on Opus 4.7/4.8). Verify current rates — via the claude-api
skill or https://platform.claude.com/docs/en/pricing — and update BASE if they've changed.