Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Coding agents

Decide if Copilot Code Review Max is worth it

Written November 2, 2026

Code Review Max runs several review agents across providers and reconciles them. GitHub estimates $10 a review at Max. Test it on one real PR first.

Audience: Maintainer deciding whether to pay for multi-agent review · Stack: A GitHub repo with pull requests, Copilot code review enabled · Tool: GitHub Copilot code review, Lite, Balanced, and Max effort levels

What changed

On 28 October 2026, GitHub put a third Copilot code review effort level, Max, into public preview, alongside Lite and Balanced. Max runs several independent review agents across models from different providers, and an aggregation agent reconciles their findings, resolves disagreements, and double-checks accuracy into a single review, according to the changelog. GitHub's code review documentation describes Max more plainly: it uses multiple high-reasoning models for the most intensive review, and carries a note that Max is in public preview and subject to change. Balanced, the default, routes a pull request to a higher-reasoning model for longer analysis. Lite gives fast, targeted feedback on common issues such as bugs, security vulnerabilities, and style inconsistencies. Max is available on Copilot Pro, Pro+, Max, Business, and Enterprise plans, per the changelog.

You choose the effort level for one pull request under Reviewers, where Copilot appears as a reviewer, or set a default in your Copilot code review settings, as the usage documentation describes. When GitHub decides which level to use, it checks the level chosen for that request first, then the level previously used on the pull request, then the level set by the person who requested the review, then the repository level, then the organization level. Administrators can also set a maximum review effort level, so Max is not available where the maximum is lower.

What it costs

GitHub publishes estimated costs per review, and every figure in this section is GitHub's estimate, not a measured price. The estimates table gives a typical range and an approximate average for each level, excluding GitHub Actions minutes: Lite $0.05 to $1, averaging about $0.25; Balanced $0.25 to $5, averaging about $1.50; Max $1.50 to $25, averaging about $10, excluding GitHub Actions minutes. GitHub states that the figures are estimates only, that consumption generally increases with pull request size and repository custom instructions, and that actual costs can fall outside the ranges and may change as models evolve.

A review has two cost components. The first is AI credits for the review itself. The second is GitHub Actions minutes for the agentic capabilities behind the review, such as gathering full project context. Using larger GitHub-hosted runners is billed at a higher per-minute rate, while self-hosted runners do not consume GitHub Actions minutes, per the documentation. The changelog adds that Max takes longer and uses more AI credits than Lite or Balanced, because it does more work on each review. By default, AI credits for an automatic review are attributed to the pull request author, and for a manual request to the person who requested it, though an organization can choose to bill the organization instead.

Strengths

Weaknesses and limits

What to do about it

  1. Run the comparison on one real pull request. Take a high-risk pull request, one where a missed issue would cost real time or real users. Request a Balanced review first and read it. Then request a Max review on the same pull request and read that. Compare the two side by side: which findings are new, which findings changed a decision you were about to make, and which findings were noise. Keep Max for pull requests where the new findings changed a decision. That comparison is the only evidence that speaks to your code; GitHub's average says nothing about it.
  2. Leave Balanced as the default. The documentation frames Balanced for security-sensitive code, multi-service pull requests, and repositories with strict quality standards, and Lite for routine changes where fast feedback matters more than exhaustive analysis. Reserve Max for high-risk or release-critical pull requests, the use GitHub names for it, and pick it per request rather than setting it as a default.
  3. Check the instructions on the head branch before a Max review. Because Max reads instructions from the branch under review, a careless instruction change in the pull request changes the review itself. Read .github/copilot-instructions.md and any path-specific instruction files in the head branch before you pay for the deepest review of a pull request that rewrote its own rules.
  4. Set the maximum where you administer repos. If you run an organization or a set of repositories, set a maximum review effort level that matches the decision you made, so Max cannot be picked casually on repositories where it was never approved. Revisit the cap when the preview status changes.
  5. Watch automatic reviews. Automatic reviews attribute their AI credits to the pull request author by default. A repository with automatic reviews on, and Max as its default, spends at the Max rate on every qualifying pull request without anyone choosing it per request. Check the automatic-review settings before you blame the per-review estimate.
  6. Let a human decide the merge. Read the Max review the way you would read the ordinary Copilot review: order the work by severity, check each suggestion against the surrounding code, and keep design decisions with a human reviewer. The aggregation agent reconciles findings; it does not decide whether the pull request solves the right problem.

Limits of this evaluation

Related playbooks

Sources

Back to: Coding playbooks hub · Also: Review a pull request with a coding agent · Copilot Fusion is generally available: what it does to your AI-credit bill · Security-review your own code with Claude Code