October 2026 Summaries

OpenAI’s Core Experimentation team describes using Codex-based GPT-6 models to automate software maintenance tasks while retaining human review and validation. GPT-6 Astra was used to investigate and remove dead code by examining repositories, pull requests, internal discussions, production profiling, and feature-flag systems, contributing to more than 1,400 merged maintenance pull requests and over 800,000 deleted lines of code, alongside smaller container images and reduced test-suite complexity. The team also applied the models to identify stale Statsig experiments and flags, including costly configurations that static analysis had missed, with one reviewed removal estimated to save 0.42% CPU in a major service. In a comparison across more than 6,000 files, GPT-6.1 Sol Ultra found nearly as many stale-experiment issues as Astra Ultra at a fraction of the API cost, although it missed some dependency-related flags, illustrating the importance of prompt refinement and repeated evaluation as models evolve. Additional uses included fixing flaky tests, improving agent instructions, and optimizing CI, with reported reductions in median CI completion time from about 10 to 4.5 minutes. The team recommends beginning with small, evidence-based draft changes, measuring results, incorporating reviewer feedback into prompts, and gradually expanding successful workflows across larger codebases.