GitHub Copilot Code Review: Lite or Balanced? Which effort fits each PR?

GitHub Copilot Code Review now offers two effort levels: Lite and Balanced. GitHub describes Lite as a fast, targeted standard review, while Balanced uses higher reasoning for a longer, deeper pass.[1][2]

The short answer: start with Lite for docs, styling, small UI fixes, and tightly scoped bugs. Pick Balanced for authorization, payments, migrations, concurrency, or changes that cross service boundaries. Pull request size alone is a poor selector. The impact of a mistake and the number of boundaries crossed tell you more.

What changes between Lite and Balanced?

GitHub documents Lite as the default for common bugs, security vulnerabilities, and style inconsistencies. Balanced routes work to a higher-reasoning model for complex logic, security-sensitive code, and cross-service changes.[1] These are effort levels inside Copilot Code Review, not two model names.

You cannot pick the underlying model separately. GitHub says Code Review uses a tuned mix of models, prompts, and system behavior, and does not support model switching.[1] If you need to choose a named model by task, that is a separate decision. The Copilot model selection guide covers it.

A task-based selection table

• README, copy, CSS, or a small visual fix: Lite.
• A tightly scoped bug in one module: Lite.
• A test or refactor that follows an existing pattern: Lite; move to Balanced if the diff spreads unexpectedly.
• Authentication, authorization, payments, or personal data: Balanced.
• Schema changes, migrations, concurrency, or cache invalidation: Balanced.
• A PR crossing services, packages, or API contracts: Balanced.
• A large but mechanical file move: start with Lite; size alone does not require Balanced.
• A tiny config change that alters production access: Balanced; fewer lines do not mean lower risk.

That table is my routing recommendation, not a GitHub benchmark. GitHub defines the intended uses but does not publish a success rate for your repository.

When is Lite enough?

Lite fits work with a narrow failure surface and an obvious check. Editing copy in one component, adding a case to an established test pattern, or fixing a bounded null check are good examples. GitHub says a requested review usually finishes in under 30 seconds.[2]

Do not read "Lite found nothing" as proof that the PR is safe. Code Review excludes some files. GitHub lists dependency management files, logs, and SVGs among them.[1] If the PR touches one of those, use a separate human or tool check.

When does Balanced earn its cost?

Balanced makes more sense when a mistake has a high impact or the change crosses several system boundaries. GitHub recommends it for security-sensitive code, multi-service pull requests, and repositories with strict quality standards.[1]

I would not treat Balanced as the "large PR" switch. A thousand-line mechanical rename may be easy to verify. An eight-line authorization change may carry real risk. Decide with three questions:

1. Which user or data could a bad change harm?
2. How many service, package, schema, or trust boundaries does it cross?
3. If tests miss the problem, how hard is production rollback?

If one answer looks ugly, starting with Balanced may be cleaner than paying for a second review later.

How large is the cost difference?

GitHub estimates that one review consumes $0.05-$1 worth of AI credits with Lite and $0.25-$5 with Balanced. Pull request size and repository instructions can raise consumption, and those ranges exclude GitHub Actions minutes.[1] These are GitHub's product estimates, not fixed-price guarantees.

Automatic review on every push means small iterations can trigger more consumption. GitHub says each review uses AI credits, while agentic context gathering can also consume Actions minutes.[1] A reasonable solo-repository default is automatic Lite for routine PRs, manual Balanced for risky ones, and a pre-merge re-review only after substantial changes.

Automatic or manual review?

GitHub's lifecycle guide positions manual review for high-context or selective use and automatic review for teams that want broad, consistent coverage.[4] Automatic review on new pushes keeps feedback current on fast-moving PRs, but it can also add cost and repeated comments to tiny iterations.

A practical solo-builder rule:

• Run Lite when each PR opens.
• Start with Balanced on PRs marked as risky.
• Re-review only when logic or security behavior changed.
• Skip automatic re-review for formatting, comments, or conflict resolution.

GitHub also warns that a re-review may repeat comments that were previously resolved or downvoted.[2] Turn on review for every push because the workflow needs it, not because the checkbox exists.

Instructions may matter more than effort

Balanced cannot repair missing context by itself. GitHub Code Review can use `.github/copilot-instructions.md`, path-specific instructions, `AGENTS.md`, and several other agent instruction files.[1][2] Keep review criteria short and testable: contracts that must not change, required checks, trust boundaries, and known architectural exceptions.

Use the AGENTS.md vs CLAUDE.md vs GEMINI.md guide to decide where persistent context belongs. Raise the effort only after the context is usable. Vague rules fed to Balanced may produce a longer wrong answer.

How should you verify the comments?

GitHub's refreshed overview groups findings into open, resolved since the last review, and previously missed. It also shows severity and links back to each inline comment.[3] That makes follow-up easier, but a Copilot comment is not evidence by itself.

Use this sequence for important findings:

1. Read the relevant call chain, not only the highlighted line.
2. Run a test or static check that reproduces the finding.
3. Read the diff before accepting a batch of suggested changes.
4. Request another review if the fix changed behavior.
5. Let human review and CI make the final decision.

The coding-agent output verification checklist gives you a fuller closing sequence. Lite versus Balanced only decides where to spend the first review budget. It does not remove the verification step.

Do not confuse effort with GPT-5.6 model choice

Code Review effort and model selection are separate controls. If you are choosing between Sol, Terra, and Luna for speed, cost, and depth, start with the GPT-5.6 Sol-Terra-Luna pillar guide. This article does not repeat that comparison. It answers the narrower PR-review question: Lite or Balanced?