Cost

init and sync spend model tokens, and so does every review. probe is read-only and costs nothing. This page explains what a run costs and how to estimate it before you start.

What costs money

gh reads cost nothing beyond the API rate limit. Codegraph runs locally and is free.

Estimate before you init

Run probe first. Its estimate block names the job count, a token range, a time range, and a dollar range, and says which basis it used:

• This repository's own recorded jobs, when at least four prior jobs exist for it. This is the closest to what the run will do.
• The calibration, otherwise, taken from the maintainers' reference runs.

If the provider's price list cannot be reached, the dollar line is omitted and the estimate says it is unavailable, rather than printing a made-up number. For OpenAI and Anthropic, which publish no price list this way, it reads estimate unavailable for this provider. The line always ends with an estimate, not a bill.

The dashboard shows the same estimate when you add a repository, and nothing starts until you confirm.

Where to watch the real number

Every run prints one summary line to stderr when it finishes:

Done in 21.0s · 3,125 in, 1,308 out tokens · $0.0016

The same numbers appear in --json under usage, and the dashboard records cost per job on the Activity page and per repository on Usage. When a price list is unavailable the line reads cost unknown rather than a false zero.

What "unknown" means

A missing cost is never shown as $0.00. Zero is what a free model actually costs, unknown is a cost that was never learned, and the two are kept apart everywhere the dashboard shows money. A single review's unknown cost is a dotted underline you can hover, or tab to, for the reason:

A total adds up only the reviews with a known cost and names how many are missing one, for example $1.23 · 4 reviews with unknown cost, rather than quietly rounding them to zero.

OpenAI, Anthropic and OpenCode never report a cost, so every review on them reads unknown. OpenRouter reports one. See AI providers.

Who pays

A remote review is paid by the server's key unless the client sends its own (see your own key). A review paid with a client's key is kept out of the server's own spend and shown after a plus sign, so a total reads $1.23 + $0.40 BYOK. If the cost of those reviews is unknown, it reads 4 reviews with unknown cost BYOK in the same place.

Keeping the cost down

--log-time prints the time each phase took, which is useful when a run feels slower than its estimate. See sync and init for the full flag list.

What is not metered

• Local gh and git reads.
• Codegraph indexing and tool calls.
• Serving the dashboard and the webhook receiver.
• Reading guides with view or view --remote.

Next: Data and privacy for where the keys and the generated files live.