Copilot code review cuts costs 20% by changing how it reads repos

AI code review now has a clearer operating metric: repository exploration is expensive, and GitHub says it found a cheaper path. Copilot code review has switched to built-in file tools from the Copilot CLI and SDK, reducing review costs by about 20% while maintaining quality.

In a June 25, 2026 changelog, GitHub says Copilot code review now uses grep, rg, glob, and view in its review path. Those tools replace custom file-exploration tooling that Copilot previously used while trying to find the relevant parts of a codebase.

The number is the interesting part. GitHub says the change, combined with instruction tuning behind the scenes, has lowered Copilot code review costs by about 20% with no drop in the review-quality standard. The company says the result has appeared in both offline and online evaluation.

The update also gives organizations more control over review depth. Teams in the Medium analysis-depth public preview can now see Medium attribution in the pull request overview comment, making it easier to confirm which review level produced the feedback. Organizations can also set a default review level for repositories that have not configured one, while individual repositories can still override the organizational default.

For engineering leaders, this is not just a small changelog item. Automated review only scales if cost, depth, and signal stay manageable across many pull requests. The move suggests that agentic developer tools are starting to optimize the boring but decisive layer: how an agent searches, opens, and reasons over files before it spends tokens writing comments.

The next question is whether the 20% saving holds in large monorepos and multi-language systems. If the same exploration stack is shared across Copilot CLI, SDK, and code review, GitHub may also reduce behavior differences between interactive coding agents and automated review runs.