CodeRabbit reported using OpenAI models to govern AI-generated code and Codex, shipping 20 times more software with an unchanged engineering headcount.
The release highlights the expanding role of automated governance as asynchronous development agents generate higher volumes of production pull requests.
As companies adopt coding agents, code reviews quickly emerge as the primary operational constraint. CodeRabbit co-founder and chief executive officer Harjot Gill stated that reviewing agent-generated work represents the exact place where engineering teams must introduce strict human judgment.
To address this scaling challenge, the platform built an AI-driven governance layer designed to validate, secure, and filter AI-produced code before deployment. Over 20,000 organizations and 150,000 open-source projects now utilize the platform to decide which automated changes are ready for production.
CodeRabbit built its initial iteration on GPT-3.5 Turbo before adopting GPT-4, and later transitioned to early access for OpenAI o1 and OpenAI o3. Because asynchronous cloud workflows run independently of local developer environments, the platform successfully harnesses models that require extended reasoning time to execute deeper, more capable analysis.
The engineering team continuously tests incoming model releases against an internal benchmark of known code issues. Gill estimated that GPT-6 Astra caught 50 percent of the issues inside that test dataset, placing it at the very top of their internal evaluations. “The nature of bugs keeps evolving, so does our benchmark as the models get better,” Gill reported.
Enhanced price-performance across OpenAI models allowed CodeRabbit to transition advanced capabilities directly from experimental research into production-ready software. Improvements in model pricing and prompt caching enabled the deployment of multiple specialized subagents throughout the platform.
Early versions of the software relied on just one or two subagents dedicated exclusively to basic code navigation. Lower-cost iterations like GPT-6 Luna now power multiple subagents across the stack, while models such as GPT-6 Astra handle complex orchestration and architectural planning tasks. “Now we are using agents everywhere on the use cases we couldn’t even afford in the past,” Gill noted.
By integrating Codex into its internal operations, CodeRabbit scaled its software output by roughly 20 times over a six-month window without expanding its engineering team. Gill personally estimated pushing between 2,500 and 3,000 pull requests in the 30-day period leading up to the announcement.
Reaching this velocity required a complete overhaul of standard software delivery pipelines, addressing bottlenecks across laptop memory, compute resources, testing suites, packaging tools, and external dependencies. The company also deployed a Codex-powered Slack bot capable of triaging production incidents, support tickets, and incoming messages, subsequently returning functional pull requests for human review.
The platform increasingly applies Codex models to non-engineering functions, managing marketing tasks and operational workflows directly through code repositories. Marketing teams handle Google Ads deployments via standard pull requests, while a dedicated content repository allows Codex to generate slide decks and customer-facing technical documents.
For application security, CodeRabbit utilized OpenAI’s Cyber model via Daybreak to execute internal penetration testing and probe its own infrastructure for complex vulnerabilities. “We found multiple vulnerabilities, advanced hacks that we could not even imagine could happen,” Gill stated.
Those testing protocols evolved into CodeRabbit Secure, a commercial offering designed to scan customer repositories. During a recent two-week promotional window covering one million dollars in free deep scans, the system scanned 13,000 repositories and flagged over 43,000 security vulnerabilities.
