Cross-model code review only helps in one direction, a July study finds

Letting one AI review another AI’s code worked. Running it the other direction made the code worse.

The experiment

A July 2026 study, “Cross-Model LLM Code Review,” ran a controlled experiment on 116 recent hard and medium LiveCodeBench tasks using Claude Opus 4.7 and Codex GPT-5.5 across six conditions.

The setup deliberately mirrored real practice: the reviewer saw the problem and the draft but could not execute tests, approximating a genuine code review step rather than an automated test run.

The asymmetry

The results do not average out to a tidy “review helps” conclusion:

• Claude Opus 4.7 reviewing Codex GPT-5.5 lifted the pass rate from 71.6 to 89.7 percent, a gain of 18.1 points
• Codex GPT-5.5 reviewing Claude Opus 4.7 pushed it down from 91.4 to 82.8 percent
• Claude reviewing its own work left 91.4 percent unchanged

One direction produced a large gain. The reverse direction produced a real loss. Self-review produced nothing at all.

What it means for your pipeline

The practical read is that a review pass is not free quality. A weaker reviewer talks a stronger writer out of correct answers: it raises objections, the writer defers, and correct code becomes incorrect code. The self-review result reinforces the point, since a model checking its own work brings no new perspective to the problem.

Direction matters more than adding another model to the pipeline. If you have wired an automatic review step into an agent chain on the assumption that more scrutiny is always better, this study is a reason to check which model sits in which seat.

The numbers here belong to two specific models on one benchmark, so they are a prompt to test rather than a law to apply. But if you run agents in a chain, review direction is now clearly a parameter worth measuring on your own tasks, not a detail to leave to whichever model happened to be convenient.