Cognition Slashes Devin Review 2.0 Costs by 82% per Pull Request

Cognition has released Devin Review 2.0 with lower prices, incremental reviews, automated fixes, and configurable specialist reviewers. The company says a typical review now costs $0.25 to $0.75, making automated review on every pull request practical for more engineering teams.

Across its customers, Cognition reports an average cost of about $2.25 for each bug found and fixed. Those economics could help teams absorb the growing volume of pull requests produced by coding agents without expanding human review capacity at the same rate.

Why review capacity lags

Coding agents can produce pull requests faster than engineers can inspect them, shifting a constraint from implementation to review. When review queues grow, teams face longer merge times, more context switching, and pressure to approve changes with less scrutiny.

Devin Review provides an automated first pass. It examines a pull request for likely defects and generates a guided tour of the changes, giving the human reviewer a map of the code before they inspect the diff.

The numbers behind the cut

Cognition attributes an 82% reduction in per-PR review costs since August to improved model routing and optimizations in its agent harness. The harness is the software layer that selects context, coordinates model calls, and manages the tools used during a review.

Incremental review accounts for part of the savings. After the first review, Devin examines only the code changed by subsequent pushes instead of processing the entire pull request again. Cognition says roughly 80% of pull requests cost less than $1 across all pushes, and teams can configure a per-PR spending cap.

The company supplied these benchmark figures, and the results may vary with repository size, language, change complexity, and review configuration. A higher finding count also requires context: teams still need to measure whether findings are accurate, useful, and distinct from issues their existing tools already catch.

Autofix routes findings back

Devin Review 2.0 can send each finding to the Devin coding agent for investigation. Devin then fixes the issue, dismisses the finding, or escalates it to a human reviewer, creating a triage loop that can resolve straightforward defects before requesting attention.

Developers can request a review by posting /devin review on GitHub or GitLab. Teams can also run reviews automatically when a pull request becomes ready, with rules based on paths, branches, labels, and draft status. Those filters can exclude low-risk or highly automated changes, such as routine dependency updates.

Two controls for risk and cost

Devin Review 2.0 adds two options for matching review effort to a repository’s risk profile:

• Devin Review Ultra: A higher-effort mode for sensitive code paths such as authentication, billing, permissions, and data handling.
• Custom Reviewers: A beta feature that accepts plain-text instructions for specialized checks beyond general defect detection. Cognition uses custom reviewers internally for backend security, database migrations, SQL performance, and interface accessibility.

Each custom reviewer has its own triggers while sharing the same Autofix workflow and spending limits. A team could therefore run a database reviewer only when migration files change, or invoke an accessibility reviewer for updates to interface components.

A crowded review layer

Devin Review competes with review features from GitHub Copilot, CodeRabbit, Graphite, Greptile, and Qodo. Cognition’s main distinction is its integration with the broader Devin agent: findings can flow into Devin Cloud or Desktop for investigation and remediation.

Teams already using Devin can add Autofix to an existing workflow with limited configuration. Other teams can adopt Review independently, using the sub-dollar per-PR pricing as a narrower entry point than assigning an entire ticket to a coding agent.

Metrics worth tracking

A useful evaluation should compare Devin Review against the team’s current process over a representative set of pull requests. The most informative measures include:

• Precision, including the share of findings developers accept or fix
• Duplicate findings already caught by linters, tests, or security scanners
• Defects found later that the automated review missed
• Time from pull-request creation to approval and merge
• Human review time saved per pull request
• Total cost by repository, change type, and review mode
• Autofix outcomes, including fixes accepted, dismissed, or escalated

The reported price reduction makes broad deployment easier to justify, while the value still depends on signal quality within each codebase. A short trial with tracked outcomes should show whether the lower review cost translates into faster merges and fewer escaped defects.