Cost

cost — Cowork → Claude API cost estimator

Overview

Claude Cowork (and Claude Code) already run on the Claude API and write one JSONL transcript per conversation under ~/.claude/projects/. Every assistant turn records the actual API call it made, including the prompt-cache breakdown. This skill reads those transcripts, sums the real token usage, and prices it at Anthropic list rates (cache-aware) so a business user can judge what their conversations would cost on the API instead of the subscription.

It answers three things:

1. Token count (current conversation and/or per conversation across history).
2. Exact tokens needed to replay the same multi-turn, tool-using conversation on the API.
3. Cost of those tokens in USD and EUR, accounting for prompt-cache hits.

When to Use This Skill

Use it when the user wants the token count or the API-equivalent cost of their Cowork / Claude Code usage — for one conversation or all of them — or asks to compare subscription vs API pricing.

How to Run

The worker is cost.py, located in this skill's folder. Run it with the system Python.

1. Pick the mode from the request:

   • current (default) — the active conversation. Use for "this conversation", "current chat".
   • history — every conversation with per-conversation + total cost. Use when the user says "all", "history", "per conversation", "each conversation", "everything".

2. Run the script with --json so you can render the result yourself:

   python "${CLAUDE_SKILL_DIR}/cost.py" --mode current --json
   

   python "${CLAUDE_SKILL_DIR}/cost.py" --mode history --json
   

   Keep the quotes — the path can contain spaces. If the command above still shows a literal ${CLAUDE_SKILL_DIR}, your surface doesn't substitute it: cost.py sits in the same folder as this SKILL.md, so locate it with Glob (**/cost/cost.py) and run the path you find.

   Options:

   • --rate-eur 0.92 — EUR per 1 USD (approximate, editable). Pass a current rate if the user gives one.
   • --include-subagents — also count subagent API calls (more complete cost). Use it when the user wants the full picture or asks about agents/subagents.

3. Render the JSON as a short English report. Show the token breakdown (uncached input / cache read / cache write / output / total) and the cost in USD and EUR. In history mode, present a ranked table (most expensive first) with a TOTAL row.

   Money formatting — always use exactly two decimals. Each cost object carries ready-to-render strings: use usd_display (e.g. $1.23) and eur_display (e.g. €1,13) verbatim. The euro figure uses a comma as the decimal separator. Do not re-derive the amounts from the numeric usd/eur fields.

Graceful Fallback (run outside Cowork / Claude Code)

If the script returns {"status": "unavailable", ...}, it found no transcripts — this happens when it's run outside Cowork or Claude Code. Do not invent numbers. Relay the message field kindly and verbatim in tone: explain that transcripts only exist inside a Cowork / Claude Code session, and invite the user to run it from there. Keep it friendly, not an error.

Reporting Notes (always include briefly)

• The figures are the actual API calls the conversation made, priced at Anthropic list rates including prompt-cache discounts (cache read ≈ 0.1×, 5-min write 1.25×, 1-hour write 2× of base input) — not an estimate.
• This is the API pay-per-token cost; the Cowork/Claude subscription bills differently, so treat it as a comparison, not your actual invoice.
• If the output lists unknown_models, mention those calls were counted as tokens but priced at $0 because their rate isn't in the table yet (add them to PRICES in cost.py).
• If the output lists estimated_models, say so explicitly — those models are newer than the price table, so they were priced with their series' most recent known rate (e.g. "claude-opus-5-1 was priced at the claude-opus-5 rate because it's newer than the table"). Their cost is included in the total, so without this caveat a fallback estimate reads as a confirmed figure.
• In Cowork, --mode history sees only the current conversation — the sandbox exposes just the active session's project folder, not the whole archive. Say so instead of presenting one conversation as the user's full history. Claude Code sees everything.
• Current rates are verified in the claude-api skill. Before adding a row to PRICES, look the rate up there rather than guessing.