对比 Copilot、Claude Code Review 与 Gemini Code Assist 的代码评审：触发、规则文件、拦截与成本

Workflows · 2026-09-24 · 1,177 words · 5 min read

How Copilot code review, Claude Code Review and Gemini Code Assist on GitHub are triggered, what they read, whether they can block a merge, and what a review costs.

code review github copilot claude code gemini

Several services now offer to review your pull requests with a model. Before switching one on, four practical questions decide whether it helps: when does it run, which files in the repository steer it, can it approve or block a merge, and what does a review cost. This post answers those four for three services, GitHub Copilot code review, Anthropic's Claude Code Review, and Gemini Code Assist on GitHub, using only what each vendor's documentation said when I read it on 24 September 2026. Where a page does not say something, this post says it does not.

When a review runs

Copilot. You can request it by hand: on a pull request, "Under 'Reviewers' in the right sidebar, next to Copilot, click Request." GitHub says a review "usually takes less than 30 seconds." Automatic reviews are set up with a branch ruleset and the rule "Automatically request Copilot code review". Two options matter. Without "Review new pushes", "Copilot will only review the pull request once", and a separate option reviews drafts.

Claude Code Review. An organization owner enables it, and each repository picks one of three triggers: "Once after PR creation", "After every push", or "Manual". A comment of @claude review starts one review in any mode, and @claude review always also subscribes that pull request to a review on every later push. Anthropic says reviews complete "in 20 minutes on average", and that a pull request from a fork is reviewed only when someone comments @claude review.

Gemini Code Assist. Reviews start on their own: "When you open the new pull request, Gemini Code Assist provides an initial review." You can ask again with /gemini review in a pull request comment, and /gemini summary posts a summary. The repository config says include_drafts defaults to true, so drafts get reviewed unless you turn that off.

The trigger is the single largest lever on cost, which the last section comes back to.

What steers the review

Each service reads its own set of files from the repository.

The detail inside each row is worth reading. GitHub's own guidance splits the job: copilot-instructions.md for "repository-wide rules specific to Copilot", AGENTS.md for "standing rules you want to share across AI tools and agents". Anthropic distinguishes its two files by strength: Code Review reads CLAUDE.md "as project context and flags newly introduced violations as nits", while REVIEW.md holds "review-only instructions" saying "what your team wants flagged, at what severity". Gemini's config.yaml has a comment_severity_threshold that defaults to MEDIUM, so low-severity findings are not posted unless you lower it, and a max_review_comments cap that defaults to unlimited.

If you already keep an AGENTS.md for coding agents (the context files post covers why), only Copilot's review documentation names it. For the other two, put the review rules in the file each one says it reads.

One exclusion is worth knowing before you rely on Gemini for security review: its documentation says it "does not generate summaries or code suggestions for any files located within the .github/workflows directory".

Can it block a merge

This is the question that decides whether the bot is an adviser or a gate, and by default all three are advisers.

Copilot: "By default, Copilot leaves a 'Comment' review, not an 'Approve' review or a 'Request changes' review," so its reviews "do not count toward required approvals". There is an exception in public preview: with Copilot approvals turned on at the repository, organization and enterprise levels, "Copilot can submit an approving review that satisfies your repository's required-approval rule the same way a teammate's approval would". A new push dismisses that approval.

Claude Code Review: findings "don't approve or block your PR, so existing review workflows stay intact". Anthropic also says "A failed run never blocks your PR".

Gemini Code Assist: the pages I read describe it posting a review with comments and adding gemini-code-assist[bot] as a reviewer. They do not say it can approve or request changes, so I cannot say either way.

The practical consequence: none of these replaces a required human approval unless you deliberately turn on Copilot's preview approvals. If you do, you are letting a model satisfy a branch-protection rule, and that is a decision for the people who own the rule. The diff-review checklist is still the human part.

What a review costs

Two of the three publish per-review figures, and they are not the same kind of number.

• Copilot: "an estimated $0.05 USD to $1 USD worth of AI credits with 'Lite' effort, and $0.25 USD to $5 USD worth of AI credits with 'Balanced' effort". These figures exclude the GitHub Actions minutes its agentic features use.
• Claude Code Review: "Each review averages $15-25 in cost", billed through usage credits separately from a plan's included usage. It is available on Team and Enterprise subscriptions.
• Gemini Code Assist on GitHub: no per-review price in the pages I read. The enterprise version lists "100+ pull requests per day" as its quota.

The listing ai-code-review-bots-compared.py turns those published figures into a monthly range. Its core is two lines of arithmetic: a review count from the trigger, times the published per-review figure.

def reviews_per_pr(trigger, pushes_after_opening):
    return 1 if trigger == "once" else 1 + pushes_after_opening

# monthly range = prs * reviews_per_pr(trigger, pushes) * (low, high) per review

For 200 pull requests a month, each pushed to three more times after opening, it gives:

Two things stand out. The trigger multiplies everything: reviewing every push quadruples the count here, and Anthropic says so directly, that after-every-push "runs on each push, multiplying cost by the number of pushes". And the services are not priced for the same job. Anthropic describes a "fleet of specialized agents" plus a verification step and a 20-minute average; GitHub says its review usually takes less than 30 seconds. A higher price for a slower, multi-agent review may be worth it or not, but these numbers alone cannot tell you, because none of the vendors publishes an accuracy rate I could compare.

If you run Claude's reviewer yourself through the claude-code-action GitHub Action instead of the managed service, Anthropic's page names two costs, GitHub Actions minutes and API tokens, and gives no per-review figure.

What the pages do not say

No vendor page I read gives a false-positive rate or a share of real bugs found, so this post cannot rank the three on quality. Gemini's pages do not state whether it can approve a pull request or what the consumer version's limits are. Claude Code Review and Copilot approvals are both labelled preview features, and GitHub adds that its cost ranges "may change as models evolve". Read the dates on the sources before relying on any figure here.

Code and data

• ai-code-review-bots-compared.py — the complete listing used in this article.

Sources

1. GitHub, "About GitHub Copilot code review" (GitHub Docs, read 2026-09-24)
2. GitHub, "Using GitHub Copilot code review" (GitHub Docs, read 2026-09-24)
3. GitHub, "Configuring code review by GitHub Copilot" (GitHub Docs, read 2026-09-24)
4. Anthropic, "Code Review" (Claude Code docs, read 2026-09-24)
5. Anthropic, "Claude Code GitHub Actions" (Claude Code docs, read 2026-09-24)
6. Google, "Review GitHub code using Gemini Code Assist" (Google Cloud docs, last updated 2026-09-22, read 2026-09-24)
7. Google, "Use Gemini Code Assist on GitHub" (Google Cloud docs, read 2026-09-24)
8. Google, "Customize Gemini Code Assist behavior in GitHub" (Google Cloud docs, read 2026-09-24)