Models are worse at reviewing their own code
Greptile
Greptile found AI coding models are worse at catching bugs in their own code than in a rival model's code. So Claude misses what Claude tends to screw up, and GPT misses what GPT tends to screw up — swapping reviewers catches more.
Based on reporting by Greptile — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Rodrigo from Greptile's research team ran an experiment that should make anyone leaning on a single AI model for code review a little nervous. He pulled 500 PRs authored by Claude Code and 500 authored by Codex, built a verified ground-truth set of roughly 1,500 real bugs, then had both Claude and GPT-5.5 review each dataset three times. The result: each model is noticeably worse at finding bugs in code it wrote itself than in code the other model wrote. GPT catches more of Claude's mistakes than Claude does, and Opus catches more of Codex's mistakes than Codex does.
The reason gets more interesting the deeper you go. Rodrigo categorized the bugs by type — performance issues, missing behavior, error handling, and so on — and found that each model's blind spots as a reviewer line up almost exactly with the mistakes it makes as an author. Claude tends to introduce more "wrong data or missing behavior" bugs, and GPT is the stronger reviewer for exactly that category. Codex leans toward speculative semantic issues and error-handling failures, and Claude is better at spotting those. Basically, a model's own coding instincts define what it considers worth worrying about, and that same blind spot follows it into review mode.
Watching the actual reasoning traces revealed a second, weirder split. Codex reviews average one or two comments per PR; Opus averages seven or eight. It's not that one model is sloppy and the other thorough — it's that GPT hunts depth-first, spending most of its context budget verifying whether a suspected bug is real, while Opus goes breadth-first, front-loading its attention on scanning the diff and flagging anything that could plausibly be wrong. Rodrigo found GPT would often identify the correct bug internally, in its reasoning trace, and then just not post it — a casualty of OpenAI's review system prompt nudging the model to minimize noise, which triggered internal tug-of-wars between
My take — AI-written commentary, not fact-checked reporting
This is the most useful piece of applied AI research I've read in a while because it turns a vague vibe —
Read more about this at: Greptile