AI Code Review Cost Per Pull Request: Seats vs Usage

Published vendor prices and launch coverage put AI code review at between about $0.25 and $25 per review, and more when a pull request gets reviewed several times. As of October 2026, GitLab lists a flat $0.25 per review and Cursor says an average Bugbot run costs $1.00–1.50. Claude Code Review was reported at launch to average $15–25 per review. Per-seat plans work the other way round: you pay a fixed monthly price, so the cost per pull request is that price divided by the number of pull requests your team opens.

Three ways AI code review is priced

Two things follow. Per-review and usage-based prices both scale with the number of reviews, so the setting that decides when reviews run (once per PR or on every push) matters as much as the unit price. Usage-based prices also scale with the diff: a large refactor costs more to review than a one-line fix. Flat per-review prices and seat prices stay the same whatever the size.

Some products combine models. GitHub Copilot code review is part of Copilot plans (Pro, Pro+, Business and Enterprise), and GitHub's changelog notes that reviews can also come from non-licensed users billed through direct org billing. Since June 1, 2026 Copilot code review also bills AI Credits and, on private repositories, GitHub Actions minutes.

Published prices, as of October 2026

These come from each vendor's own pages or from launch coverage. Several changed during 2026, so check them again before you set a budget.

Each one in more detail:

• Claude Code Review. Anthropic's costs page says the service has no flat per-review price. The cost of a review scales with the review level (Quick, Standard or Deep), the size of the change, the complexity of the surrounding code and how many candidate issues need verifying. The $15–25 average was reported by Help Net Security at the March 2026 launch. The organization pays, not the developer who opened the pull request. It is available on Team and Enterprise plans.
• GitHub Copilot code review. GitHub's changelog says reviews on private repositories draw Actions minutes from the plan's included amount and bill any overage at standard Actions rates. On public repositories, Actions minutes remain free. Self-hosted and larger runners are billed at different rates. GitHub's docs add that Balanced reviews use more AI credits and may use slightly more Actions minutes.
• Cursor Bugbot. Cursor announced in May 2026 that Bugbot is moving to usage billing. Existing customers switch at their first renewal after June 8, 2026. Effort levels (Low, Default, High, Smart) trade cost against depth. A review that is skipped because the pull request is too large records no usage.
• GitLab Code Review Flow. GitLab charges the same flat rate for every merge request and project. It is available on GitLab Duo Agent Platform on GitLab.com, Dedicated, and self-managed 18.8.4 or later.

Cost per pull request, worked through

Per review or usage:  monthly cost = PRs per month × reviews per PR × average cost per review
Per seat:             monthly cost = billed seats × seat price
Either model:         cost per PR  = monthly cost ÷ PRs per month

The example below uses assumed numbers, not benchmarks: 10 developers, 200 pull requests a month, 3 pushes per pull request.

Things to keep in mind when reading the table:

• These are review charges only. Each product also needs its base plan: Claude Team or Enterprise, a Cursor plan, a Copilot plan or Copilot org billing, or GitLab Duo Agent Platform.
• Anthropic says reviewing on every push multiplies cost by the number of pushes. It also says a later automatic review focuses on what changed since the last one, so your repeat-review costs may not match the launch average.
• Cursor's and GitLab's pages don't say whether every push starts another billed run. Their every-push figures assume it does, so treat them as upper bounds.
• The Warpway figure assumes all 10 developers authored a reviewed pull request in the last 30 days. Prices exclude taxes.

To find where a seat price breaks even against a per-review price, divide one by the other. At $30 a seat, that is 20–30 Bugbot runs or 120 GitLab reviews per developer per month. This compares prices, not what you get for them. The tools differ in whether they block merges, ask people questions, or only post comments.

The traps that push the bill up

Reviewing on every push

In Claude Code Review, a repository you add from the admin page starts on After every push, which is the setting that runs the most reviews. Switching to Once after PR creation or Manual cuts the number of reviews. @claude review always subscribes a single pull request to reviews on each push, and @claude review stop ends that subscription. In After every push mode, a comment can't stop push reviews. Converting the pull request to a draft does. If a push leaves the pull request adding and removing the same lines (an empty commit, a clean rebase, or a clean merge of the base branch), Code Review can keep the earlier result instead of running again.

Copilot reviews a pull request once unless Review new pushes is turned on. With per-push billing, every fixup commit and every push that answers a review comment can start another billed review.

Spend caps that skip reviews

Claude Code Review skips reviews in three cases: the Code Review spend limit is used up, the organization's overall spend limit is used up, or the usage credit balance runs out. It also skips once the remaining budget is less than the expected cost of one review. The limit resets at the next billing period, so if you hit it mid-month, pull requests opened for the rest of the month go unreviewed unless someone raises it. Developers see a comment on the pull request. A Claude review never approves or blocks a pull request, so a skipped review doesn't stop the pull request from merging.

Decide ahead of time what a skipped review should do. If it has to stop a merge, the skip needs to show up in a required check. On Warpway's Free plan, a private pull request over the 20-review monthly limit is not reviewed. Warpway / Review reports Usage limit reached instead of a result, and Warpway never reports success for a review it did not do. The Team plan, which includes merge gating, has no fixed monthly cap.

Large diffs and deeper review levels

With usage pricing, the most expensive reviews are the large, complex pull requests. Claude's Deep level, Bugbot's High effort and Copilot's Balanced reviews all cost more than the lighter settings. Cursor reports that in its internal runs, High effort found 35% more bugs. Claude lets an Owner set the level per repository, for example Quick for routine code and Deep for critical code. That is one way to spend the budget where the risk is.

Two meters for one review

Copilot code review on a private repository uses AI Credits and Actions minutes at the same time, and the runner type changes the minute rate. GitHub's changelog recommends checking your Actions budget and monitoring both Copilot and Actions usage. Check which billing option (member or organization) your reviews use, because that decides what happens when credits run out.

How Warpway's plans, usage limits and seats work

Warpway is a review orchestrator for GitHub pull requests, installed as a GitHub App. These figures come from the Billing doc:

Usage limit (Free). Each private pull request counts once per UTC calendar month. New commits and reruns of that pull request in the same month don't use another review, and a review refused because of the limit doesn't count. Settings → Billing shows how many reviews you have used.

Seats (Team). A seat is a person who authored a pull request that Warpway reviewed in the last 30 days, in any of the organization's repositories. Bots are not counted, and the minimum is one seat. Warpway updates the seat count from this activity, and Stripe prorates the change. People who only answer Warpway's questions don't take a seat, and people who answer in Slack don't need Warpway accounts.

Pushes. Warpway starts a new review on each push, and lenses whose files didn't change keep their previous result. On Team, the number of reviews per pull request doesn't affect the bill, because billing is by seat.

Trial and downgrades. Each organization can take the 14-day Team trial once, before its first subscription, and no card is needed to start it. If the trial ends without a payment method, or a subscription is cancelled or ends after failed payments, the organization moves to Free and nothing is deleted. On Free, trust levels are capped at Assist, so the check reports neutral and no longer blocks merges. Private reviews already run that month, including those run on Team, count toward the Free limit. The Enterprise plan (SSO, bring-your-own-key, a dedicated tenant) is planned but not yet available.

Estimate your own cost before you commit

1. Count pull requests per month, per repository:

   gh pr list --state all --search "created:2026-09-01..2026-09-30" \
     --limit 1000 --json number --jq 'length'

2. Estimate pushes per pull request. Commits per merged pull request is a rough proxy: it overcounts multi-commit pushes and undercounts force-pushed branches.

   gh pr list --state merged --limit 200 --json commits \
     --jq '[.[].commits | length] | add / length'

3. Count distinct authors to compare against seat pricing, then subtract bot accounts:

   gh pr list --state all --search "created:2026-09-01..2026-09-30" \
     --limit 1000 --json author --jq '[.[].author.login] | unique | length'

4. Multiply by the trigger setting you will actually use, not the default.

5. Run a pilot of any usage-priced tool with a spend cap. Anthropic suggests a few representative repositories set to Once after PR creation or Manual, then reading each repository's Average cost column.

6. Write down what happens when the cap is hit, and who will notice.