Code review and code search
Martian's Code Review Bench is the strongest independent board: open code, 200k+ real pull requests, scoring whether developers act on a tool's comments, updated daily. Its leader keeps changing, so any single ranking here would be stale within months.
Independent benchmarkMartian Code Review Bench (via dev.to and aicodereview.io summaries) · summaries read Nov 7, 2026
Who led Martian's Code Review Bench
- Mar 25, 2026cubic65.7% F1
- Mar 2026CodeRabbit#1, later #5 by July
- Jul 30, 2026Greptile60.8% F1, 76.2% precision
- After Jul 30, 2026Qodo#1; the source gives no date
Vendor-reported / commissionedSignal65, commissioned analyst study (commissioner undisclosed) · Mar 2026
Signal65 bug-detection study: CodeRabbit strongest overall
93
true positives, CodeRabbit
4
false positives, CodeRabbit
96%
precision, CodeRabbit
Tools
Alphabetical. Links go to each vendor's own site.
| Tool | What it does | Evidence | Price or terms |
|---|---|---|---|
| CodeRabbit | PR reviewer; GitHub, GitLab, Bitbucket, Azure DevOps | none yet | $24/user/mo annual; free for open source |
| Cursor Bugbot | PR review agent, low false-positive focus | none yet | with Cursor plans |
| Greptile | whole-codebase-context PR review + API | none yet | free starter; $1/credit, no seat fees |
| Qodo (PR-Agent) | PR review + tests; 13.2k-star open-source core | none yet | open-source core free; hosted paid |
| searchcode | code search API and MCP | none yet | free tier + API key |
| Sourcebot | self-hosted code search with an agent API | none yet | open-source core + hosted |
Caveats
- Five organisations published code-review benchmarks in six months and each crowned a different winner (Git AutoReview).
- Greptile reports an 82% catch rate on its own benchmark; Augment's re-run on the same five repos found 45%.
- Academic 2026 benchmarks (SWE-PRBench, CR-Bench and others) score models, not commercial review tools. On SWE-PRBench, frontier models catch 15–31% of human-flagged issues from the diff alone.