Summary

  • GitHub says Copilot code review uses AI credits for model tokens and GitHub Actions minutes for agentic infrastructure. Its published AI-credit estimates for Lite, Balanced and Max exclude the Actions component; the review model is chosen automatically and is not disclosed.
  • Actions minutes attach to the repository and its billing owner. AI-credit attribution follows the requester, pull-request author or organization policy. Standard hosted runners in public repositories and self-hosted runners are free on the Actions meter; included allowances can also absorb eligible use.

The review comment has two accountants

A pull request receives a Copilot review. The developer sees comments on code, perhaps a suggestion worth keeping. The organization sees two different consumption trails. GitHub charges the model interaction in AI credits. For the agentic machinery that gathers project context and can pass suggestions onward, it records GitHub Actions minutes. One visible review is two operational purchases with different units and, often, different owners.

GitHub's pricing page puts the split in plain sight. The model behind a given code review is selected automatically and not disclosed, so the exact token price can vary. The published ranges for Lite, Balanced and Max are estimates of AI-credit consumption only; they expressly leave out Actions minutes. Adding a quoted AI-credit estimate to an Actions-minute count would produce a neat looking number with no unit. The actual cash effect depends on plan allowances, runner type and budgets.

The attribution split is where the market story sharpens. Actions minutes are assigned to the repository, then to its enterprise or cost center where applicable. AI credits can be charged to the member or organization under the chosen policy. For automatic pull-request reviews, the author is generally the credit source; a manual request generally follows the requester. Bot and unlicensed paths can move that cost to the organization. The person who turned on automatic reviews, the person who opened the PR and the owner of the repository need not be the same budget holder.

The runner choice adds another lever. Copilot's agentic review uses standard GitHub-hosted runners by default. Larger runners cost more per minute; self-hosted runners do not consume GitHub Actions minutes. Standard hosted runners in public repositories are free on that meter. If Actions is unavailable, GitHub says the review can still appear, with fewer agentic capabilities. A review completed without its context-gathering worker is a different product experience, and a different cost trace, from one that ran the full path.

An owner can reconcile the two trails instead of debating an imaginary flat price per review. GitHub identifies dynamic/agents/copilot-pull-request-reviewer as the billing-report workflow path for Actions usage. Pair that record with the AI-credit report, the review effort setting, runner choice and the automatic-review rule that triggered the job. The control question is simple: who approved the review, and whose allowance did each part spend? A comment in a PR does not answer it.

Sources