Skip to content

cost.md

Latest commit

 

History

History
48 lines (39 loc) · 2.2 KB

cost.md

File metadata and controls

48 lines (39 loc) · 2.2 KB
Raw

Cost

Each model call carries the whole Claude Code system prompt, including every skill description you have installed. That, not the length of your request, is the price of a check. On a subscription it comes out of your usage limits; with an API key it is billed per token. A few dozen normal runs can take a noticeable share of a five-hour limit.

Measured numbers

All with sonnet.

What Cost
lint, check, list, init free
one batch call, demo suite (6 skills, 12 cases) $0.05–0.10
one batch call, about 80 skills installed $0.14
one normal run, demo suite about $0.05 per case

A normal run stops Claude Code as soon as it loads a skill or starts answering, so you pay for the routing decision and not for the work.

Spending less

  • Go up the levels. lint on every change, run --batch when lint is clean, a normal run only for what the batch flagged. The batch report ends with the --only line to confirm its failures.
  • Run only what you touched. --skill a,b keeps the cases that mention those skills.
  • Set a budget. --budget 2 stops starting new runs once the spend estimate reaches $2. A run can end without reporting its cost, for example when it hangs and is killed; skillcheck then prices it from the tokens it used, at the rate of the runs that reported one (or at list price before any has). The report's cost line then starts with ~, and --json keeps the exact part in costUsd and the estimate in estimatedCostUsd.
  • Keep repeat: 1 while iterating. --repeat 3 makes a result stable and triples the price. Use it in CI or when a case looks flaky.
  • Fewer installed skills make every call cheaper. In CI, point --config-dir at a directory with only the skills under test (see ci.md).

Why the batch is cheap and why it is not the final word

run --batch puts up to 25 requests into one prompt and asks the model for a structured answer: which skill would it load for each. The model sees its real skill list, but it states a choice instead of making one, and the two can differ. On the demo they matched every time, but treat a batch failure as a lead and a batch pass as a good sign, not a proof.