OpenAI Maintains Software with GPT-6: 1,400+ PRs, 800k Lines, 92%-for-14%

Marketing copy about OpenAI's models is everywhere; a first-person account by an OpenAI engineer of using those models to maintain the company's own software is rare. Maintaining software with the GPT-6 family appeared on the developers.openai.com blog with a page date of Oct 14, 2026, written by an engineer on the Core Experimentation team — the group that "builds and maintains the software that supports experiments across the company." This site fetched the full text on 2026-10-18; every number below comes from the post, and all of it is officially self-reported.

Dead code: a typo-laden one-liner replaces homegrown static analysis

Act one is dead code. The author recalls the pre-agentic setup — complex static-analysis systems wired into runtime logging — and the incidents when removals went wrong anyway. The turning point came in the GPT-5.6 Sol era: targeted prompts with clear validation criteria let the model write the static analysis and remove dead code itself, at similar reliability and lower maintenance cost. GPT-6 Astra lowered the bar further: one short prompt, typos and all, explicitly inviting the model to "check slack, statsig ... or any other source you need."

Astra's evidence trail is the most informative part of the post: it read past PRs, Slack threads, and internal project trackers, and inspected production performance profiling tables to confirm the deleted lines were never reached — explaining its rationale for each deletion. Five engineers then ran variations of the prompt across repositories, merging more than 1,400 PRs and deleting over 800k lines part-time over two weeks. The officially reported gains: one service's container image shrank from 3.31GB to 1.72GB (48%), and another cleanup removed about 400 visual snapshot variants from the test suite.

Stale experiments: having Astra write its own audit prompt

Act two is harder — stale experiments are not necessarily unused, and a wrong removal breaks production. The author brainstormed with Astra and had the model draft a /goal audit prompt: scan the whole repository for Statsig gates, configs, experiments, and layers (including indirect usages through enums, constants, wrappers, aliases, schemas, dynamically assembled names, and cross-service consumers), reconcile production-referenced constants, preserve killswitches, and "do not stop after producing a conservative shortlist."

The result: hundreds of experiments and feature flags retired. The showpiece case shows the model's judgment — a single line of Python from years ago mapped an internal override flag to all other flags, hiding among hundreds of similar mappings and causing significant CPU fan-out. Astra quantified the cost but escalated for human review because the flag was explicitly marked "Permanent" in the management console; after review it was removed, with an estimated 0.42% CPU saving in the main affected service.

The model comparison: 92% of the findings at 14% of the cost

Act three is cost. The author ran the same audit on three model/effort combinations, over subdirectories covering 6,000+ files and 1 million lines of Python, TypeScript, Go, and Rust, validating the findings by hand:

The official framing: GPT-6.1 Sol Ultra produced 92% as many findings as GPT-6 Astra Ultra at 14% of the API cost; Medium is cheaper still but finds less. Costs use standard API pricing and include research subagents and the follow-up turns Codex ran automatically. The discrepancy pattern is named too: on Ultra, most of the gap came from constant feature flags used as dependencies in other flags — Astra identified those, 6.1 Sol missed them — and OpenAI recommends folding such patterns into prompts while warning that future model releases can make the corrections obsolete.

CI speedups, and OpenAI's own caveat

The post closes by pointing the same method at CI: a prompt that ranks load-reduction ideas by estimated daily CI minutes saved led to about 30 optimization PRs; median CI completion fell from roughly 10 minutes to 4.5, and the 90th percentile from 12.5 to 6. One standalone fix — adding a missing mock — cut a 63-test resolver suite from 592.6 seconds to 29.8 seconds, a 95% reduction with all 63 tests retained.

The post declines to mythologize: some proposals, it says, added more complexity than their expected benefit justified, and prioritization and measurement remained an important part of the workflow — in larger codebases the models find more optimizations than a team can absorb, and instrumentation is what kept that from becoming noise.

Next steps

• Teams wanting to replicate: the three /goal prompts (flaky-test fixes, AGENTS.md instruction audits, CI load-reduction rankings) are copyable from the post; OpenAI suggests starting with GPT-6.1 Sol Medium and raising effort if results are promising.
• Tracking model positioning: this site's The GPT-6 family grows: gpt-6-sol, gpt-6-luna, gpt-6.1-sol land in the official SDKs holds the type-level facts on the four model IDs in the SDK enums; this post adds their official self-reported cost profile on a real task.
• Citation note: every figure here is OpenAI's own (including internal estimates like the 0.42% CPU saving); third parties cannot audit them, so cite with that framing.