Copilot code review effort levels went generally available on 7 August. There are two of them, they got new names, and the level a review ran at is now visible on the pull request itself.
That last part is the one that changes anything. A setting nobody can see is a setting nobody can debug.
The two levels
GitHub frames them by the kind of change rather than by how much compute they burn:
- Lite — straightforward, routine changes that need focused feedback
- Balanced — complex, sensitive or cross-service changes that need deeper analysis
The rename, and why it is not just cosmetic
Low and Medium became Lite and Balanced. Existing configurations moved across automatically, so nothing breaks.
The old names implied a ladder with a missing rung: if there is Low and Medium, where is High? Users reasonably assumed they were being denied the good one. Lite and Balanced describe a choice between two fits rather than two rungs, which matches what the feature actually is.
It also lines up with the direction the rest of the stack is moving. Reasoning effort became a setting you have to think about across these tools this summer — Copilot's cloud agent got its own reasoning-level control on 3 August. Two vocabularies for the same idea in the same product would have been a mess.
Per review, or set it once
You can pick the level when you request an individual review, which is the right control for a person who knows this particular change is hairy.
Administrators can also set an organization-wide default under Organization Settings → Copilot → Copilot code review. Repositories inherit that default unless they set their own preference, which is the sensible hierarchy: a policy for the common case, an override for the repository where the common case does not apply.
It is available on Copilot Pro, Pro+, Max, Business and Enterprise.
The visibility change is the real feature
The level used now appears in timeline events and in the pull request overview comment. That sounds like a detail and it is the thing that makes the rest usable.
Before this, "Copilot reviewed it" was a single undifferentiated claim. A reviewer looking at a merged PR could not tell whether the machine had given it a glance or a proper read, so the sensible response was to trust it uniformly — which means trusting the light reviews too much and the deep ones too little.
Now the record shows the effort level, so a human reading the pull request later knows what kind of check actually happened. That is what makes it possible to say "anything touching payments gets Balanced" and then verify that it did.
A sensible default policy
Do not set the whole organization to Balanced because deeper sounds better. Deeper reviews on trivial changes produce noise, and reviewers who learn to skim Copilot comments stop reading the ones that matter.
Lite as the organization default, Balanced inherited by the repositories where a mistake is expensive — auth, payments, infrastructure, anything with a migration — is the policy that matches how the levels are described.