OpenAI engineers cut 800,000 lines of code using Codex and GPT-6 Astra

OpenAI published a blog post on October 18, 2026, describing how it automated internal software maintenance using its Codex coding agent and models from the GPT-6 family, digitaltoday.co.kr reported.

OpenAI gave GPT-6 Astra short, vague tasks to find and clean up unused code, drawing on past pull requests, Slack conversations, internal project records and production performance data. Five engineers merged more than 1,400 pull requests over two weeks alongside their other work, removing more than 800,000 lines of code. For one service, the container image size fell from 3.31GB to 1.72GB, a 48% reduction.

Astra was also used to clean up old experiments. It identified CPU-heavy waste, leading the team to remove hundreds of experiments and feature flags. In one case it flagged a flag set to "permanent" as CPU-heavy; after human review the team deleted it, cutting CPU usage on key services by about 0.42%.

Cheaper model tested

The team also tried the cheaper GPT-6.1 Sol on 6,000 files and 1 million lines of code. At the highest reasoning setting it found 131 of the 142 issues Astra found, about 92%, at an API cost of $15.12 versus $109.06 for Astra, roughly 14%.

Applying about 30 pull requests based on Codex suggestions cut median CI completion time from about 10 minutes to 4.5 minutes. A batch of 63 tests dropped from 592.6 seconds to 29.8 seconds, a 95% reduction.

OpenAI advises starting with small maintenance tasks, reviewing the results, and then expanding the scope.