Greptile

Sources#

What it is#

Greptile sells an AI code-review agent that comments on pull requests. It is not a model lab: the main review agent is "usually a frontier model from either OpenAI or Anthropic," so the product is the harness, the repository context and the routing, not the reviewer's weights. That architecture is what makes its research possible — swapping the underlying model is a configuration change, which is the manipulation its central study runs.

What it published#

"Models are worse at reviewing their own code" (Rodrigo Caridad, research team, 2026-07-21) — two 500-PR ground-truth datasets, one Claude Code-authored and one Codex-authored, ~1,500 verified high-severity bug comments. Each frontier model catches fewer bugs in its own family's code than in the other's: Opus 4.7 53.7% same-model vs 60.0% cross-model, GPT 5.5 50.5% vs 62.0%. See Same-Model Review Blindness.
Model Inversion — the product feature built on that finding: detect the authoring agent from the PR's trail, route the review to the other vendor's model.
A State of AI coding report and a body of blog explainers on AI code review, which circulate in practitioner discourse (see 3100 Opinions on Code Review in an AI World: Building Causal Theory from Practitioner Discourse) alongside recurring "Greptile alternatives?" threads — the tool has enough adoption to generate both citations and churn complaints.

It also appears as a measured object rather than an author in Tran et al.'s secret-leak study: Greptile is one of seven bots and human reviewers that commented on any of the 74 genuine live credentials, the pool whose combined comment rate is the 18.9% floor on Security Debt of Agent-Generated Code.

How to read it (evidence posture)#

Greptile's flagship study is filed case-study, corrected down from the raw document's empirical at compile — the standing [evidence] lint warning on that row is the correction, not drift. The measurement is real work, but the ground truth is vendor-built from "sentiment analysis, upvote/downvote ratios, and git archaeology" with no released artifact, no labelling protocol, no inter-rater agreement, no validation of the LLM judge doing the matching, and no confidence interval on gaps of 6-12 points. One arm's prompt was tuned against the outcome metric until recall recovered. And the conflict of interest is direct: cross-vendor routing is exactly what a single-model reviewer cannot offer, so the finding sells the product built from it. The subject measured is two third-party models run through their own vendors' review features, which is why the tier is case-study rather than vendor-claim.

Trust the direction; attribute magnitudes to Greptile; the discriminating replication — a third model family across both corpora — is a standing open question on Same-Model Review Blindness.

Connections#

Same-Model Review Blindness — the finding it measured and named, and the page that carries the correction of its evidence tier
Review as the Control Point — its study supplies the automated-reviewer-capability moderator a lineage variable that theory had assumed away
Optimizer–Evaluator Decoupling — the maker/checker rule its data first tests on model lineage rather than on role
Security Debt of Agent-Generated Code — where its bot appears as one of the seven commenting actors behind the 18.9% detection floor
Claude Opus 4.7, Codex — the two authoring agents whose corpora the study is built from