The team says GPT-6 Astra gathered evidence from past pull requests, Slack, project trackers, Statsig, and production profiling before preparing changes for review.
Five engineers ran cleanup work part-time over two weeks and merged more than 1,400 pull requests, deleting over 800,000 lines of code. OpenAI reports that one service image fell from 3.31 GB to 1.72 GB, while another cleanup removed about 400 visual snapshot variants. For stale experiments, the agent produced reports rather than changing live services; one subsequently reviewed removal was associated with roughly 0.42% CPU savings on a service.
In one 6,000-file, roughly one-million-line audit, OpenAI reports 142 findings for GPT-6 Astra Ultra at $109.06 and 131 for GPT-6.1 Sol Ultra at $15.12. It also says approximately 30 CI optimization pull requests reduced median CI time from about 10 to 4.5 minutes. The post is an internal case study and does not establish that these results, costs, or safety outcomes will reproduce in other repositories.