claude-cost

claude-cost

Price your local Claude Code sessions at the Anthropic API rate card.

Adds an /api-cost slash command that walks ~/.claude/projects/*.jsonl, computes a shadow cost for each assistant turn from its usage metadata, and prints two tables:

1. Top N most expensive sessions - with turns, input/output tokens, cache hit %, and the dominant model, so you can see why a session was expensive, not just that it was.
2. Cost by repo - tells a one-off spike apart from a whole-repo pattern.

Costs are shadow prices against the public Anthropic API rate card. They represent the dollar value a Max subscription extracts, not dollars actually paid.

Requirements

• Claude Code
• uv to run the script via its PEP 723 header
• Python 3.11+ (the script uses only the standard library)

Install

claude plugin install destilabs/claude-cost

Or clone this repo and point Claude Code at it as a local plugin.

Usage

/api-cost                       # last 30 days, top 10 sessions
/api-cost --days 7              # last 7 days
/api-cost --since 2026-04-01    # since a specific date
/api-cost --top 25              # wider session table

The bare /cost slash is reserved by Claude Code (aliased to /usage for subscription info), which is why this plugin uses /api-cost.

What you will see

Window: 2026-04-17 to 2026-04-24
Parsed 14,423 assistant turns across 60 sessions in 12 repos.

Top 5 most expensive sessions
#  Date         Repo         Model   Turns  Input  Output  Cache   $ Cost  Session
-  -----------  -----------  ------  -----  -----  ------  -----  -------  --------
1  04-22 17:41  stitch-mono  opus-4  1,835   3.3K    1.5M    98%  $824.90  794f6f7a
2  04-20 08:10  stitch-mono  opus-4  1,855  11.4K    1.0M    98%  $729.40  a3e67dda
...

Cost by repo
Repo                 Sessions   Turns    $ Total  $ / session
-------------------  --------  ------  ---------  -----------
meta                       32   5,307  $2,111.88       $66.00
stitch-mono                 5   4,351  $1,782.80      $356.56
...
TOTAL                      60  14,423  $5,505.22       $91.75

How the pricing works

Each assistant turn in a Claude Code session JSONL carries a usage block:

{
  "input_tokens": 42,
  "output_tokens": 1240,
  "cache_creation_input_tokens": 18500,
  "cache_read_input_tokens": 95000
}

The script multiplies each token count by the per-model rate in the public Anthropic pricing page and sums. cache_creation_input_tokens is billed between 1.25x and 2x the input rate depending on TTL; the script uses the 5-minute rate, which slightly undercounts 1-hour cache writes.

Opus 4.5, 4.6, and 4.7 are billed at the lower tier ($5 input, $25 output, $0.50 cache read per million tokens). Opus 4 and 4.1 remain at the higher tier ($15 / $75 / $1.50). The script encodes both. Unknown models fall back to the current Opus tier (conservative over-count for new models).

Scope

v1 is deliberately minimal. Not included:

• Model-mix or slash-command breakdowns
• Week-over-week trends or time-series charts
• HTML / JSON export
• Max-subscription ROI framing
• Alerts or budget pacing

These may come later if there is demand. Open an issue.

License

MIT