CodeRabbit claims 20x more output, without defining the yardstick

CodeRabbit says in an OpenAI customer story that Codex helped it ship roughly 20 times more software than six months ago with the same engineering headcount. Co-founder Harjot Gill estimates that he submitted 2,500 to 3,000 pull requests in the 30 days before the interview.

Codex increased production while CodeRabbit took control

The team uses more agents for implementation, incidents and support. CodeRabbit can verify and ship low-risk changes automatically, while people retain approval over sensitive work. The company also uses GPT 6 Luna for cheaper subtasks and GPT 6 Astra for planning.

The bottleneck moved from writing to deciding

More generated code creates more changes that someone must understand, test and approve. CodeRabbit is demonstrating Codex while selling the control layer needed to handle its output. For engineering teams, the workflow shift matters most: people write less and decide more often what may pass.

The 20x figure does not measure delivered value

Neither OpenAI nor CodeRabbit disclosed a release baseline, change sizes, defect rates or customer outcomes. The number is company-reported and cannot show that another team would reproduce it. Thousands of small pull requests are also not the same as 20 times more value.

Fixes and human intervention will reveal the quality

Useful evidence would include the share of automatically merged changes, post-deployment regressions, false positives and human review time. Those measures will show whether agents shortened the route to a finished product or merely delivered a larger queue for inspection.

Lilith's verdict

Agents dropped 3,000 pull requests onto the conveyor. The person at the end of the line still decides how many belong in the customer’s box.