Gitar, Sonar's code verification agent, now uses Claude Opus 5.5 for code review, so teams get useful feedback faster. Our testing shows Opus 5.5 cut median review time by about half versus Opus 5, with higher recall and precision.
Measured improvements
Metric
Improvement
What changed
Higher precision
+7% points
A larger share of findings matched the expected issues.
Higher recall
+2% points
A larger share of the expected issues was identified.
Less time to review
50%
Lower median time to complete a review session.
Token efficiency
65%
Lower blended model cost per token.
Code review beyond the diff
Opus 5.5 works with Gitar's cross-repository context and memory of team guidance. For Enterprise teams, that means reviews can account for code in other repositories and explanations from earlier review conversations.
Cross-Repo Analysis checks whether a change breaks a contract another repository depends on. An API change in a service can be checked against its clients. Findings identify the affected repository and which side needs a fix.
Knowledge gives reviews memory. If a team explains that generated files must be changed through their schema, Gitar can carry that guidance into later reviews. It selects the knowledge relevant to the changed code, with citations back to the source.
Teams can inspect that knowledge and mark outdated guidance stale. The context behind a review can evolve with the codebase, without requiring the same explanation in every thread.
More investigation in fewer steps
Opus 5.5 completed reviews with about half as many tool-call turns and over 10× fewer subagents. It combined related code inspections and handled more of the investigation directly.
Subagents still helped with focused questions, such as checking an external library's behavior. The model used them selectively while working through local code itself.
We also noticed more reviews relying on Bash alone for direct inspection. Dedicated reading and search tools still helped: keeping them alongside shell commands produced stronger review results than a Bash-only setup in our testing.
Try Opus 5.5 in Gitar
Point Gitar at your next pull request and watch a faster, sharper review with Opus 5.5 land on your own diff. Don't just review. Ship.
Try Gitar*, or learn more about Gitar AI code verification.
* Teams using their own keys for models may see availability on a different schedule.
