CodingSoftware & developer tools
Greptile uses GPT-6 Astra's computer use to run and test the apps changed by pull requests
Greptile says it uses GPT-6 Astra's computer-use ability to launch the app each pull request changes and test it like a user. The aim is to catch visual and interaction bugs that diff-only review can miss. It also uses Codex and OpenAI's Daybreak.
From the OpenAI News article of October 26, 2026 and cited sources: For each pull request, Greptile starts a dev server and has GPT-6 Astra click through the user flows and check the screen visually. [1] [2]
SOURCES[1] OpenAI: Greptile takes bug detection beyond code review… openai.com[2] Newsroom America coverage newsroomamerica.com
AT Pulse research deskPublished · Updated · 4 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- For each pull request, Greptile starts a dev server and has GPT-6 Astra click through the user flows and check the screen visually. [1] [2]
- Original source
- OpenAI News article, company source
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- October 26, 2026 article published
- Covered
- Oct 27 edition
- Evidence
- 9 sources across 9 sites
- 1OpenAI: Greptile takes bug detection beyond code review…openai.com
- 2Newsroom America coveragenewsroomamerica.com
- 3StartupHub.ai: Greptile's live app tests need receiptsstartuphub.ai
- 4The Hacker News: OpenAI Launches Daybreakthehackernews.com
- 5Wikipedia: GPT-6 Astraen.wikipedia.org
- 6Blockchain.News: Greptile says GPT-6 Astra extends bug…blockchain.news
Reach: not measured yet
For each pull request, Greptile starts a dev server and has GPT-6 Astra click through the user flows and check the screen visually. In the main example, the model shrank the browser to phone size and found that a button had disappeared. [1] [2]
Runtime testing is added on top of Greptile's existing static review. That review builds a graph of the repository and runs parallel agents on each pull request. Greptile already had sandboxed pull-request testing through its earlier TREX runner, so this was not its first runtime testing. [6] [9]
Greptile's Knowledge Base indexes the codebase and serves it to Codex through an MCP server. CEO Daksh Gupta says giving agents this context reduces reverts. For security, Greptile uses Daybreak, OpenAI's tightly controlled program that confirms vulnerabilities in a sandbox and proposes patches for human review. [1] [2] [4] [7]
The evidence is one OpenAI-published customer story with a single phone-viewport example and no independent measurement.
Read the full assessment
If it holds up, AI code review is moving toward running the changed app, which overlaps with QA and likely costs more compute per pull request.
Executive brief
OpenAI published a customer story on October 26, 2026 about Greptile, an AI code-review startup. In it, Greptile says it uses GPT-6 Astra's computer-use ability to run the app a pull request changes and click through it like a user would. It does not only read the diff. The example given: the model shrank the browser to iPhone size and noticed that a button had disappeared (Newsroom America). The story also covers Greptile's use of Daybreak for security and Codex for fewer rework cycles. It gives no catch rates, false-positive rates or cost figures (StartupHub.ai).
What changed and event timeline
OpenAI launches Daybreak
A security program that combines GPT-5.5 variants, the Codex Security agent and partners including Cisco, CrowdStrike and Palo Alto Networks. Access is tightly controlled ().
GPT-6 Astra goes public
A limited preview opened September 3. The model targets multi-step work, including computer use and software engineering ().
Greptile adds Plus and Apex tiers
These are more intensive review tiers for pull-request analysis ().
Reported $50M Series B
Led by Menlo Ventures at a $750M valuation, according to one outlet. No second source confirms this ().
OpenAI publishes the Greptile story
It describes GPT-6 Astra for workflow testing, Daybreak for security and Codex for reducing rework ().
Capabilities and access
- GPT-6 Astra: Greptile starts a dev server for each pull request, resizes the browser to a phone viewport, steps through user flows and checks the screen visually (Newsroom America).
- Daybreak: Builds threat models, confirms vulnerabilities in a sandbox and proposes patches for human review. Access is by request or through sales (Help Net Security; The Hacker News).
- Codex: Connects to Greptile's Knowledge Base through an MCP (Model Context Protocol) server.
Read the full section
- GPT-6 Astra: Greptile starts a dev server for each pull request, resizes the browser to a phone viewport, steps through user flows and checks the screen visually (Newsroom America).
- Daybreak: Builds threat models, confirms vulnerabilities in a sandbox and proposes patches for human review. Access is by request or through sales (Help Net Security; The Hacker News).
- Codex: Connects to Greptile's Knowledge Base through an MCP (Model Context Protocol) server.
- The story does not name exact API model IDs or pricing.
Technical analysis for researchers and developers
- Architecture: Greptile builds a graph of a repository's files, functions and dependencies, then runs parallel agents over each pull request (Blockchain.News).
- Context delivery: The Knowledge Base indexes the codebase and writes documentation for humans and agents.
- Evaluation: The story documents no method for measuring results: no catch rate, miss rate, flake rate or definition of a "revert" (StartupHub.ai).
Read the full section
- Architecture: Greptile builds a graph of a repository's files, functions and dependencies, then runs parallel agents over each pull request (Blockchain.News). Computer use adds a runtime layer on top of that static review.
- Context delivery: The Knowledge Base indexes the codebase and writes documentation for humans and agents. Codex reads it through MCP, which means any MCP-capable agent could consume it.
- Evaluation: The story documents no method for measuring results: no catch rate, miss rate, flake rate or definition of a "revert" (StartupHub.ai). The results cannot be reproduced from public material.
Claims and evidence
- Astra catches visual and interaction bugs that diff review misses. The only support is a single phone-viewport demo (Newsroom America).
- Giving agents context means fewer reverts. This is CEO Daksh Gupta's statement and comes with no data.
- Codex Security has helped fix more than 3,000 critical and high-severity vulnerabilities. This figure comes from OpenAI, as summarized by Reco.
Read the full section
- Astra catches visual and interaction bugs that diff review misses. The only support is a single phone-viewport demo (Newsroom America).
- Giving agents context means fewer reverts. This is CEO Daksh Gupta's statement and comes with no data. The story says "thousands of customers" have connected Codex to the Knowledge Base, which is the only number in it.
- Codex Security has helped fix more than 3,000 critical and high-severity vulnerabilities. This figure comes from OpenAI, as summarized by Reco.
- Independent corroboration: none found. All of the claims come from an OpenAI-published customer story.
Context and prior work
- Greptile was founded in 2023 by Daksh Gupta, Soohoon Choi and Vaishant Kameswaran and went through Y Combinator's Winter 2024 batch (Blockchain.News).
- It earlier raised a $25M Series A led by Benchmark. Runtime testing therefore existed before the Astra integration.
- Codex Security launched in March 2026 and is the agent that does the work inside Daybreak (Reco).
Read the full section
- Greptile was founded in 2023 by Daksh Gupta, Soohoon Choi and Vaishant Kameswaran and went through Y Combinator's Winter 2024 batch (Blockchain.News).
- It earlier raised a $25M Series A led by Benchmark. Greptile v4 shipped March 5, 2026, and the TREX sandboxed pull-request test runner shipped in June 2026, according to a company-profile page (Yespress). Runtime testing therefore existed before the Astra integration.
- Codex Security launched in March 2026 and is the agent that does the work inside Daybreak (Reco).
Limitations, safety and contested findings
- StartupHub.ai argues that claims of speed mean little without a false-positive rate, a catch rate or a before-and-after revert curve.
- Blockchain.News also notes there are no benchmarks or customer comments (Blockchain.News).
- Reports say GPT-6 uses a "recurrent depth" technique that could hide the model's reasoning.
Read the full section
- StartupHub.ai argues that claims of speed mean little without a false-positive rate, a catch rate or a before-and-after revert curve. It also flags that Astra's cost and latency per pull request are not disclosed (StartupHub.ai).
- Blockchain.News also notes there are no benchmarks or customer comments (Blockchain.News).
- Reports say GPT-6 uses a "recurrent depth" technique that could hide the model's reasoning. Safety researchers have raised concerns about how well that reasoning can be monitored (Wikipedia).
- Daybreak access is restricted, so most teams cannot reproduce Greptile's security setup (The Hacker News).
Business and practitioner implications
- AI code review is moving toward running and testing the changed app, which competes with parts of traditional QA.
- Driving a browser with a computer-use model costs more compute than reading a diff.
- Exposing codebase context over MCP lowers switching costs between coding agents, and it also makes Greptile's Knowledge Base a place customers can get locked in.
Read the full section
- AI code review is moving toward running and testing the changed app, which competes with parts of traditional QA. Before replacing existing tests, ask vendors for measured catch, miss and flake rates (StartupHub.ai).
- Driving a browser with a computer-use model costs more compute than reading a diff. Budget for that per pull request.
- Exposing codebase context over MCP lowers switching costs between coding agents, and it also makes Greptile's Knowledge Base a place customers can get locked in.
- Greptile is relying on OpenAI for workflow testing, security and coding all at once, which concentrates vendor risk.