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

Step	Spends tokens	Why
`probe`	No	Reads the repository through `gh` only
`init`	Yes	Extraction and synthesis jobs over code, pull requests, and history
`sync`	Yes, but less	Reuses unchanged evidence from cache, so only changed inputs rebuild
Large-file diff summary	Yes, small	A diff over 500 changed lines is summarized with the low model instead of being sent whole
Local and PR `review`	Yes	Runs on your OpenRouter or Hetzner key
Remote `review`	On the server	The server's key pays, not the laptop's
GitHub App auto-review	On the server	Same as remote

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. 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.

Keeping the cost down

Lever	Effect
`--max-pr-months`, `--max-commits`, `--max-pull-request-change-lines`	Fewer inputs means fewer extraction jobs
`--only-request-changed-pr`	Keeps only pull requests with a `CHANGES_REQUESTED` review
`--improve-matrix=1`	The default. A higher value adds improvement passes, each another full call
`sync` instead of `init`	Reuses unchanged rows, so a moving default branch costs less
Low model for extraction	`--low-model` does the history work. The high model handles synthesis and review
`--disable-codegraph`	Does not change token cost, but shortens the run

--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.