Cross-model review is asymmetric: Claude reviewing Codex helps
Controlled pairings of Claude and Codex as writer/reviewer show one direction lifts pass rates; the reverse can hurt.
Source: Cross-Model LLM Code Review: Should you use Claude to review Codex or vice versa?
This paper runs the experiment teams argue about in Slack: should model A review model B’s draft? Six conditions approximate a practitioner workflow—solo baselines, both cross-model orders, both same-model orders. The reviewer sees the problem and draft but cannot run tests.
Claude reviewing Codex drafts lifts pass rate from 71.6% to 89.7%. Codex self-review also helps. The reverse does not: Codex reviewing Claude drops 91.4% to 82.8%, and Claude self-review is flat. Useful pairing is asymmetric.
Why it matters for code review
“Ensemble review” is not automatically good. Model identity and role assignment matter. If you wire multi-model PR review, measure directionality; a weak reviewer on a strong draft is not a free safety layer.