TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Greptile says it is using GPT-6 Astra’s computer-use capabilities to test software changes by launching applications, changing browser sizes and simulating user interactions. The approach extends its bug-detection work beyond reading code, though the source does not provide independent performance results or details on general availability.
Greptile says it is extending AI-assisted code review into application testing with GPT-6 Astra, using the model’s computer-use capabilities to open a development version of a pull request, interact with it in a browser and inspect what appears on screen. The company says this can expose bugs that are difficult to find by reading code alone, including interface problems on smaller screens.
In an account published by OpenAI, Greptile co-founder and CEO Daksh Gupta describes a workflow in which an agent starts a development server for a proposed code change, sets a browser to an iPhone-sized viewport and simulates a user moving through the application. GPT-6 Astra can then use its visual understanding to inspect the interface. Gupta’s example is a button that disappears when a website is viewed at a mobile screen size.
The aim is to test behavior as well as code. A pull request may look logically sound in a code review but still fail when the application runs, when a user follows a particular path or when the layout changes. Greptile presents browser-based checks as a way to add those forms of validation to its agents’ review and testing work. The report does not describe the workflow as a replacement for human review or as a guarantee that every defect will be found.
OpenAI’s report also describes other parts of Greptile’s work, including Daybreak, which Gupta says is intended to help customers strengthen security, and a company Knowledge Base that indexes codebases and makes documentation available to Codex through an MCP server. Gupta says thousands of customers have connected Codex to that knowledge base. These are separate elements of the company’s broader effort to provide coding agents with codebase context and validation tools.
Testing Changes in a Running App
The approach addresses a gap between reviewing a proposed change as text and checking what users actually experience. Code inspection can miss runtime and visual defects, especially problems that appear only under a particular screen size or sequence of interactions. Running the application gives an agent another source of evidence to use when assessing a change.
That matters as teams use AI tools to produce and review software more quickly. Gupta said that relying on a conventional QA firm to test every change would be impractically slow relative to the pace at which developers can ship code with these models. Automating some browser checks could help teams add testing without requiring a manual tester to repeat every interaction. Whether it reduces defects in practice will depend on how reliably the system can launch applications, choose useful test paths and recognize failures.
For software teams, the practical distinction is that the described system does not stop at suggesting whether code looks correct. It attempts to observe the result of a change in a live interface. That could make automated review more useful for applications with substantial user-interface behavior, but the report provides no measured detection rate, error rate or customer outcome data for the GPT-6 Astra testing workflow.
Products Worth Considering
Logitech MX Mechanical Wireless Illuminated Keyboard, Tactile Quiet, Graphite
Jackery HomePower 3600 Plus Portable Power Station, 3584Wh, 3600W
Logitech MX Master 4 for Mac, Ergonomic Bluetooth Mouse with Haptic Feedback, Ultra-Fast Scrolling, USB-C Charging, for MacOS, iPadOS, Windo
As an affiliate, we earn on qualifying purchases.
From Code Review to Browser Checks
Greptile builds agents that review and test code changes using information about the wider codebase, according to the company description in OpenAI’s report. Its stated goal, as Gupta put it, is “reducing bugs and making the world’s software more reliable.” The GPT-6 Astra example reflects a broader idea in that work: validating a change may require running the software, rather than relying only on a static reading of the patch.
The report also links Greptile’s workflow to its codebase Knowledge Base. The company indexes customer code and documents how it works, then makes that information available to Codex through a Model Context Protocol server. Gupta said customers use that context so Codex can write code that follows existing patterns, and that the company tracks how often AI-generated code is reverted. He said providing agents with context leads to fewer reverts, but the report gives no figures or methodology for that comparison.
OpenAI’s account is a company-published description of Greptile’s use of OpenAI products. It offers an example of the intended workflow, not an independent evaluation of GPT-6 Astra’s testing performance. Details needed to compare it with conventional automated testing or other computer-use systems are not included in the supplied material.
“We are in the business of reducing bugs and making the world’s software more reliable.”
— Daksh Gupta, Greptile co-founder and CEO
Performance Data Not Provided
The OpenAI report does not say when the GPT-6 Astra workflow will be generally available, which Greptile customers can access it, or whether it is already deployed beyond the described use case. It also does not provide test results showing how often the system catches bugs, how often it produces false alarms, or how its results compare with existing browser automation and human QA.
The mobile button example is presented by Gupta to explain the workflow; the report does not identify it as a benchmark result. Details are also missing on how agents select user journeys, handle authentication or application setup, and report a suspected defect to developers. Those operational details would help teams judge the tool’s coverage and the human oversight it needs.
Other claims in the report, including the statement that thousands of customers connected Codex to Greptile’s Knowledge Base and that contextualized agents produce fewer reverts, are attributed to Gupta. No customer list, baseline, time window or underlying measurements are supplied. The report also does not provide technical detail on Daybreak’s security methods, so its safety and effectiveness cannot be independently assessed from this material.
Availability and Validation Details
The supplied report does not announce a launch date, pricing, a customer rollout plan or a named next milestone for GPT-6 Astra-powered testing. Greptile’s next public details would need to clarify access, supported applications and the steps an agent can test before teams can assess how the feature fits into existing review pipelines.
Evidence about performance would also help establish the tool’s value: for example, reported results across real pull requests, the kinds of bugs detected, and how often human reviewers confirm an agent’s findings. Until Greptile or OpenAI releases such information, the confirmed development is the described capability and workflow, while its reliability and practical impact remain unquantified.
Gupta also said Greptile sees faster development as a goal and described OpenAI’s Ultrafast as a way to accelerate work. The report does not give a schedule for that effort. For now, the immediate development is Greptile’s use of GPT-6 Astra to bring browser-based checks into its broader code-review and testing approach.
Key Questions
What is Greptile doing with GPT-6 Astra?
Greptile says it uses GPT-6 Astra’s computer-use and visual capabilities to run a development version of a pull request in a browser, interact with the application and inspect the interface for potential problems.
How could this find bugs that code review misses?
Some defects appear only when software runs or when a user takes a particular action. Greptile’s example is a button that disappears at an iPhone-sized browser viewport, a visual issue that may not be apparent from reading the code alone.
Has Greptile shown that the system improves bug detection?
The OpenAI report describes the intended workflow but gives no detection rates, false-alarm rates or independent evaluation. Its effectiveness cannot be determined from the supplied material.
When will customers be able to use the feature?
The report does not specify a release date, availability terms or which customers can access GPT-6 Astra-powered application testing.
What other products or tools does the report discuss?
It also describes Greptile’s Daybreak security work and its Knowledge Base, which the company makes available to Codex through a Model Context Protocol server. Those initiatives are discussed separately from the browser-testing example.
Source: openai.com
