Devin Review 2.0: automated review on every PR
Cognition announced Devin Review 2.0, now the default PR review tool when working with Devin. The company says the cost of reviewing a PR dropped 82% since August, with a typical review costing $0.25-$0.75. The release adds incremental reviews, auto-review rules and a beta of Custom Reviewers.
Key points
- Review cost per PR is down 82% since August, with a typical review at $0.25-$0.75
- Findings per PR rose from 1.22 to 1.32 while median latency fell from 139s to 94s
- `/devin review` in GitHub or GitLab starts a one-off review; rules automate it like CI
- Re-runs are incremental by default, and a per-PR spend limit caps total cost
- Devin Review Ultra trades higher cost for a more thorough pass on sensitive code
- Auto-review rules can filter by draft status, file path, branch and label to skip low-value PRs
Cognition shipped Devin Review 2.0, now the default PR review tool when working with Devin. The company says reviewing a PR got 82% cheaper since August, with a typical review at $0.25-$0.75 and about 80% of PRs staying under a dollar including every push. Findings per PR rose from 1.22 to 1.32 and median latency fell from 139 to 94 seconds.
Reviews start with a `/devin review` comment in GitHub or GitLab, or from Devin Cloud and Desktop, and can run automatically like CI. Rules filter out draft PRs, file paths, branches and labels; re-runs are incremental by default, and a per-PR spend limit caps cost. An Ultra effort level buys a deeper pass, and beta Custom Reviewers run your own checks.
Real-world use
- Running Ultra effort on PRs that touch billing or authentication code where a missed bug is expensive
- Adding CI-style automatic review on every PR in a small team so humans start with context and findings
- Writing Custom Reviewers to enforce safe database migration and SQL performance rules on every PR
- Excluding low-value PRs in a busy monorepo by filtering on file paths and labels
Why it matters
At under a dollar per PR, reviewing everything rather than a curated subset becomes affordable, so agents can take the first pass and humans review findings instead of hunting for them.