OpenAI says five engineers, working alongside other duties for roughly two weeks, merged more than 1,400 Codex maintenance proposals and cleaned up over 800,000 lines of code. Astra cross-checked PR history, internal conversations, and production profiles, and requested human review for permanent configuration flags. One service image shrank 48%, while median CI time fell from about ten minutes to 4.5. In a separate comparison, Sol Ultra found about 131 of Astra’s 142 findings at 14% of the API cost, but also missed dependent flags. As a company-reported case, it supports starting small and checking evidence and regression tests before expanding.