Greptile, the AI code review startup, now goes beyond reading diffs. A new case study from OpenAI says the company uses GPT-6 Astra to run the apps it reviews. Astra clicks through them like a user would and catches bugs that never show up in the code.
Astra is one of three OpenAI products in the setup. Greptile pairs Astra’s computer use with Daybreak, OpenAI’s cybersecurity program, and with Codex, OpenAI’s coding agent. Here’s what each one does.
1. Astra tests the app, not just the code
Code review tools have a blind spot. They read the changes in a pull request, but they don’t see what the user sees. A layout bug that hides a button on small screens can look fine in a diff.
OpenAI’s example shows how Greptile fills that gap. Astra starts a development server for the pull request, shrinks the browser to an iPhone-sized screen and moves through the app like a real user. Because it can read the screen, it spots the button that disappears on mobile.
That’s a real shift. The reviewer stops asking “does this code look right?” and starts asking “does this feature work?”
2. Daybreak backs up security
Greptile also uses Daybreak to help its customers build stronger defenses. Astra went to Daybreak members first when it launched on September 3. OpenAI says Astra reaches the “Critical” cybersecurity level under its Preparedness Framework, its internal system for rating how risky a model’s abilities are. That’s why its strongest security features go through a program that checks who gets access.
The case study doesn’t say much here. The security work comes up in a quote from Greptile CEO Daksh Gupta, but the case study doesn’t explain how Daybreak fits into the review process.
3. Codex gets the context it needs
The third piece is about cutting rework. Greptile’s Knowledge Base writes up how a company’s codebase works, for both people and AI agents, and makes those docs available to Codex.
The logic is simple. Coding agents guess wrong less often when they know the team’s conventions, structure and past decisions. And fewer wrong guesses means fewer pull requests that have to be redone.
What it means for dev teams
Teams already using Greptile can expect three practical wins:
- Fewer UI bugs in production. Visual and user-flow problems get caught before code is merged.
- Less manual QA. An agent does the click-through checks a person used to do.
- Better agent output. Codex works from written-down context instead of guessing.
What stands out is where this is heading. AI code review started out as a smarter linter, the kind of tool that flags style and syntax problems. Now it’s turning into an automated QA tester that runs your app. Other code review tools will probably feel pressure to add live testing too.
Access and caveats
GPT-6 Astra rolled out in stages. Daybreak members got it first, followed by ChatGPT Plus, Pro, Business and Enterprise, the API (as gpt-6-astra) and Amazon Web Services. Enterprise admins have to switch it on. The case study doesn’t cover what Greptile charges for these features.
A few things are worth keeping in mind:
- It’s a vendor case study. OpenAI published it, so it’s promotional by nature.
- No hard numbers. Greptile doesn’t share how many bugs it catches, how many false alarms it raises, how often changes get rolled back or what any of it costs.
- Computing power isn’t free. Starting a dev server and running a browser agent on every pull request costs far more than reading a diff. Teams will want to know if the extra bugs caught are worth the bill.
What comes next
If live app testing works at scale, code review could soon mean checking whether a change actually works for the user, not just whether the code looks right. Greptile is betting it will. The next thing to watch is whether the company publishes real data to back that up. You can read the full case study on OpenAI’s site.