GitHub launches ReviewBench, an open AI code-review agent benchmark built from 219 real pull requests across 19 languages

What changed

GitHub launched ReviewBench on October 5 as an offline benchmark for AI code-review agents. GitHub analyzed distributions across 103.9 million pull requests, then selected 219 pull requests from 187 public open-source repositories spanning 19 languages. Its golden set combines human review, follow-up fixes, static analysis and frontier-model findings under a common rubric, with independent senior-engineer relabeling reporting 96.6% agreement. GitHub is publishing the dataset, methodology, judge prompt/configuration and a self-serve runner for third-party agents and leaderboard submissions.