How good is GitHub Copilot code review? Pricing and results

How good is GitHub Copilot code review? Our results

Eight coding agents, built on Claude, GPT, Grok, DeepSeek and GLM models, built the same two features in an open-source analytics app. HeyDeer and Copilot code review reviewed all 16 pull requests, and we checked every finding against the code.

HeyDeer found more issues on the pull requests of all eight agents, with the same number of false positives. Costs are pay-as-you-go prices for the credits each review used, without plan allowances. The full benchmark has every finding and the dataset.

On ReviewBench, GitHub’s open benchmark for AI code review, three Copilot code review tiers sit between 93% and 97% precision in the default high-severity view. HeyDeer’s precision there is 99.0%, the highest on the board.

How GitHub Copilot code review works

1. Request a review. Open a pull request, then choose Copilot under Reviewers. From the terminal, run gh pr edit PR-NUMBER --add-reviewer @copilot.
2. Or review automatically. Turn on automatic review in personal settings, or add a branch ruleset with “Automatically request Copilot code review,” optionally including new pushes and drafts.
3. Give it your rules. Copilot reads .github/copilot-instructions.md, path-specific .github/instructions/*.instructions.md files and AGENTS.md.
4. Apply or hand off fixes. Comments include suggested changes you can apply in a couple of clicks, and Fix with Copilot hands them to the Copilot cloud agent, which opens a pull request against your branch or commits to the same one.

Copilot skips dependency files such as package.json and Gemfile.lock, log files and SVG files. Beyond GitHub.com it runs in the GitHub CLI and Mobile, VS Code, Visual Studio, Xcode and JetBrains IDEs, and in Azure DevOps in public preview.

What a Copilot review looks like, and how to act on it

On GitHub.com a review usually arrives in under 30 seconds. Copilot posts an overview comment plus inline comments on the lines it’s worried about, and labels each comment High, Medium or Low severity so you can triage the list from the top.

By default the review is a plain “Comment” review: it never approves or requests changes, and it doesn’t count toward the approvals your branch protection requires. With Copilot approvals turned on, an approving review from Copilot can satisfy that rule. If new commits land after Copilot approves, the approval is dismissed.

Working through the comments takes three moves:

• Accept a suggestion. Commit one suggested change on its own, or batch several into a single commit. In VS Code, Apply and Go To Next edits your working copy without committing.
• Delegate the fix. Fix with Copilot opens a draft comment where you tell the cloud agent which feedback to address, then pick a new pull request or a commit on the current one. It needs both code review and the cloud agent turned on.
• Rate the comment. Thumbs up or down is how you give Copilot feedback. A written reply is for your teammates: Copilot can’t see it and won’t answer.

Two habits save time. Copilot reviews a pull request once and doesn’t re-review new pushes on its own, so use the re-request button next to its name in Reviewers after you fix things. And expect repeats: a re-review can raise a comment again even after you’ve resolved or downvoted it.

Writing custom instructions Copilot follows

Copilot reads its instructions from the pull request’s head branch, so a change to them can be tested in the same pull request that introduces it. Besides .github/copilot-instructions.md, path-specific instruction files and AGENTS.md, it also picks up CLAUDE.md, GEMINI.md and REVIEW.md.

GitHub’s own advice: start with 10 to 20 specific rules aimed at your most common review problems, write them as short imperatives rather than paragraphs, and keep any file under about 1,000 lines, since shorter files are more likely to be fully processed. A repository-wide file might look like this:

# Code review instructions

## Security
- Flag hardcoded secrets, API keys or tokens.
- Every query on tenant data must filter by organization_id.

## Data
- Store money as integer cents, never floats.
- New database columns need a migration and a default.

## Tests
- A bug fix needs a test that fails without the fix.

Rules that only apply to some files go in .github/instructions/*.instructions.md with an applyTo pattern in the frontmatter, such as applyTo: "**/*.py", which keeps Copilot from applying Python rules to your TypeScript.

Some requests are ignored. Copilot code review doesn’t support instructions that change the format of its comments, add emoji, edit the overview comment, block merging, or follow external links; paste the content you’d link to into the file instead. Vague requests such as “be more accurate” don’t help either. Instructions also make reviews longer: GitHub notes that consumption grows with pull request size and with your custom instructions.

Turning on automatic review for an organization

Personal settings only cover your own pull requests. To review every pull request across many repositories, an organization uses a ruleset:

1. In the organization’s Settings, open Repository, then Rulesets, and create a branch ruleset.
2. Under Target repositories, include or exclude repositories by name pattern, such as *feature. Under Target branches, add the default branch or all branches.
3. Under Branch rules, select Automatically request Copilot code review. Add Review new pushes to get a fresh review on every push, and Review draft pull requests for drafts.

A developer’s own settings can’t switch off push or draft reviews that a ruleset turns on. The organization also decides who pays: with member billing, the default, each review draws on the author’s own credits and fails once they run out; with organization billing, the organization pays, which first requires enabling paid AI Credits usage. Business and Enterprise organizations can also let Copilot review pull requests from members without a Copilot license, if an owner or administrator enables it.

GitHub Copilot code review pricing

Since June 1, 2026, Copilot bills usage in GitHub AI Credits at $0.01 each instead of premium requests. Each plan includes a monthly credit allowance; Business and Enterprise overage is charged per credit. Code review also consumes GitHub Actions minutes, except on self-hosted runners.

GitHub’s average cost per review is $0.25 with Lite effort, $1.50 with Balanced and $10.00 with Max, which is in public preview. Those are averages: GitHub gives typical ranges of $0.05 to $1 for Lite, $0.25 to $5 for Balanced and $1.50 to $25 for Max, with larger pull requests at the top. Lite focuses on common issues such as bugs and security problems; Balanced, the default, sends the pull request to a higher-reasoning model; Max uses several of them.

What the same workload costs

Take 10 developers who each open 10 pull requests a month: 100 reviews. At the Balanced average, those reviews use about 15,000 AI Credits. Ten Business seats cost $190 a month and include 19,000 credits, so reviews alone use most of the allowance the team also needs for Copilot’s other features. The same workload costs $14 a month on HeyDeer, with no seats.

Copilot code review vs HeyDeer, side by side

When Copilot is enough, and when to add another reviewer

Copilot is enough when every developer already has a seat with credits to spare, pull requests are small and written mostly by people, and someone reviews each change after it. In that setup it’s a fast first pass that costs nothing extra to try.

Add a second reviewer when coding agents write a large share of your pull requests, when a missed bug is expensive, or when review credits start crowding out the rest of the allowance. GitHub itself says Copilot isn’t guaranteed to spot every problem and recommends human review alongside it. An independent reviewer that runs on the same pull requests shows quickly whether Copilot is missing issues on your code.

• Copilot code review fits teams where every developer already has a Copilot seat, reviews happen mostly inside GitHub and the IDE, and a quick first pass is enough.
• HeyDeer fits teams that ship code from several agents and people, want measured precision and more issues caught before merge, and prefer one workspace price to a seat per developer.

Still deciding? Run both on the same pull requests with our evaluation guide, or see all AI code review tools ranked by precision.

Sources

1. GitHub Docs: About Copilot code review
2. GitHub Docs: Copilot plans
3. GitHub blog: Copilot is moving to usage-based billing
4. GitHub Docs: Using Copilot code review
5. GitHub Docs: Configuring automatic code review
6. GitHub Docs: Using custom instructions to unlock the power of Copilot code review