claudeworth

Is your Claude subscription actually worth it? claudeworth compares your real Claude Code usage against pay-as-you-go API rates and tells you the effective discount you're getting.

$ npx claudeworth

That's it. It reads ~/.claude/projects/ locally, prices every message at Anthropic's published API rates, and opens a browser dashboard.

What it shows

• Effective multiplier — your API-equivalent value ÷ your prorated subscription cost (e.g. "7.4× this week").
• Lifetime return — total API-equivalent value extracted vs. subscription actually paid since your first session, as a net dollar figure and a multiplier. Steadier than the single-week number.
• Last 5 hours — proxy for Claude Code's per-session cap.
• This week — value vs. break-even, with a monthly pace projection.
• By project — which repos are eating your subscription (named from each session's real working directory).
• By model — Opus vs Sonnet vs Haiku split, with per-model cache hit rate.
• Top sessions — your most expensive sessions, with the model mix that drove the cost.
• Heatmap — day-of-week × hour-of-day, toggleable between cost and call count, with each day's total down the side. Find your peak hours.
• All-time stats — lifetime totals, current streak, most expensive single call.
• Methodology — the exact formula, the pricing table, and a list of any model strings that didn't match (collapsible).

Why another one?

There are already two tools in this space: codeburn and toktrack. Both are good at "here's what your usage cost." Neither answers the question subscribers actually have: am I getting my money's worth, or should I switch to API billing?

claudeworth is built around that one question. Everything on the dashboard ties back to it.

It also fixes one small but expensive bug both other tools appear to share: Opus 4.5/4.6/4.7 quietly dropped from $15/$75 per MTok to $5/$25 per MTok in late 2025. Tools that hardcoded the old rates over-report cost by ~3× on Opus-heavy users.

Privacy

Nothing leaves your machine. No network calls. No telemetry. No accounts. No "sync to cloud."

Your Claude Code session files contain every prompt and response you've ever sent. A tool that ships them anywhere is a privacy disaster. claudeworth reads them, computes locally, writes an HTML file, opens your browser. That's the entire data flow.

You can verify this by reading the source — it's small (6 files, ~1,200 lines of vanilla JS, no dependencies).

Usage

# Open the dashboard (defaults to Max 5x; switch plans with the dropdown in the page)
npx claudeworth

# Set the initial plan via CLI (you can still change it in the dropdown)
npx claudeworth --plan pro     # Claude Pro    ($20/mo)
npx claudeworth --plan max5    # Claude Max 5x ($100/mo)  [default]
npx claudeworth --plan max20   # Claude Max 20x ($200/mo)

# Just write the HTML, don't open the browser
npx claudeworth --no-open --out ./dashboard.html

The plan dropdown at the top of the page lets you compare what your same usage would be worth on each tier — useful for deciding whether to upgrade, downgrade, or stay put.

Requirements

• Node.js ≥ 18
• Claude Code installed and used at least once on this machine (so ~/.claude/projects/ exists)

How the value calculation works

For every assistant message in your session logs, claudeworth reads the usage block and applies Anthropic's published rates:

cost = input_tokens          × base_input_rate
     + output_tokens         × base_output_rate
     + cache_read_tokens     × base_input_rate × 0.1
     + cache_5m_write_tokens × base_input_rate × 1.25
     + cache_1h_write_tokens × base_input_rate × 2.0

Cache multipliers and per-model rates: see Anthropic's pricing docs.

Sum it all up, divide by your prorated weekly subscription cost ($23.01/wk for Max 5x), and you have your effective multiplier.

There is no "% of allowance" gauge because Anthropic doesn't publish hard token quotas for the subscription tiers — the limits are described in fuzzy terms ("~225 Sonnet msgs per 5h" for Pro, multiples for Max). claudeworth deliberately doesn't fabricate a percentage; it just shows API-equivalent value.

Known limitations

• Heuristic 5m/1h cache split: when a message's cache_creation block lacks the 5m/1h breakdown, all cache-create tokens are treated as 5m writes (the default). Slightly under-counts 1h-cache-heavy workloads.
• Local timezone: days are bucketed in your machine's local timezone. If you work across timezones, the daily/heatmap views may look off.
• Session cap is a rolling 5h window, not a true Claude Code session anchored at the first message after a 5h gap. Close approximation, not exact.
• Subscription tier limits aren't enforced — see methodology above.

Contributing / reporting issues

Open an issue or PR on the repo. Particularly useful: model strings that show up under "unknown" in the methodology section.

License

MIT