Maintaining software with the GPT 6 family | OpenAI Developers
OpenAI's Core Experimentation team details how GPT-6 Astra and GPT-6.1 Sol transformed software maintenance at scale—accelerating code cleanup, rooting out stale experiments, and making review cycles more efficient and cost-effective. The piece argues that with the right prompts and agentic context, these models not only rival traditional static analysis and scripts for reliability, but let human engineers focus on higher-order control and judgment. The team presents real numbers on cleanup volume, cost savings, and model comparison, and recommends prompts others can adapt for similar gains.
Transcript
Vince You ever wonder what happens when you hand your entire codebase over to a coding agent and just say, 'Go clean this up'? Because apparently, that's now a normal week at OpenAI.
Ava I mean, if your baseline is 'deep static analysis plus human review', it's wild how fast agentic approaches have caught up. The post claims five engineers using GPT-6 Astra deleted over eight hundred thousand lines—in fourteen hundred pull requests. And that was mostly a side project.
Vince It's so much cleanup, it's almost suspicious. The scale, I mean—not the method. But the argument here is they actually got better reliability than their old static analysis pipelines, because Astra could cross-reference past PRs, Slack threads, production metrics—context you just can't jam into a rules engine.
Ava Right. And the model didn't just yank things blindly—it explained why each removal was safe, which made reviews way faster. For me, that's the difference: it's not just code deletion, it's receipts. You're not rubber-stamping a diff, you're reacting to the model's evidence.
Vince Okay, so here’s the part that leapt out—one container shrank from three point three gigs to one point seven after the deletion campaign. Forty-eight percent less bloat, just from dead code and test artifacts. That’s not some theoretical win—that’s production fat gone.
Ava Sure, and the stale experiment stuff is even knottier. Static analysis just misses things when eligibility or config gates move around. But Astra builds its own prompt, runs a full audit, and even investigates CPU cost of old flags. That anecdote about the permanent feature flag eating CPU cycles? It's subtle, and the system spotted it but still flagged it for human review since it was marked permanent.
Vince Yeah, and the human review made the call to drop it, saving almost half a percent of CPU in a hot codepath. It’s those details that sell the approach for me—it’s not just bulk deletion for show, it’s actual engineering judgment.
Ava Mm-hm.
Vince And did you catch the API cost numbers? Six point one Sol did ninety-two percent of Astra Ultra's findings at fourteen percent of the API price—fifteen bucks versus one hundred and nine. It’s not just smarter, it’s cheaper and repeatable.
Ava Yeah, that's real. And the article’s honest about the gaps: Sol missed some dependencies that Astra caught, especially when one feature flag depended on another. But their feedback loop is tight—they tweak prompts as they go, and Sol closes the gap.
Vince Right, and the workflow is now basically: write an ambiguous prompt, watch the model self-refine it, vet the results, and keep humans in the loop for edge cases. They even have Astra suggesting which prompts to run next.
Ava It does raise the bar for what 'code hygiene' means. If your agent can cross-reference Slack, PRs, production telemetry, and even propose new cleanup tasks, the old static analysis feels kind of—
Vince —like a relic, yeah. But I have a soft spot for boring-control-panel infrastructure. You know I'm going to say it: this is exactly that boring-but-works thesis. The real win here is the control layer—prompt design, receipts, human gatekeeping.
Ava Yeah… okay, that's fair.
Vince And honestly, for a quick read like this, we’re trusting the source’s numbers more than re-deriving every experiment. So if one of those cost figures is off by a buck or two, caveat emptor. I’d want the receipts for a proper deep-dive.
Ava Right. But the broad shape feels right: coding agents as routine maintenance workers, not magic wizards. It’s just fast, explainable labor. And if you don't have to buy a mega-rig for static analysis anymore—just spin up a prompt and review the receipts—that changes who can do this work.
Vince I love that the 'just try it' prompts survive all the way to production. The author even lists their favorites for flaky test cleanup and agent context audits. I can see teams adapting those, especially with Six point one Sol pricing.
Ava Yep. And the fact that they closed the loop between model findings, human review, and cost tracking? That’s the part most teams still miss when they 'adopt AI.'
Vince The thing I can't decide is whether this makes full hands-off codebase maintenance feel closer, or if it just raises the bar for what hands-on means. Like, are we getting more autonomous, or just more finely supervised?
Ava Honestly, both. It’s more autonomous at the detection and summarization layer, but now your review workload is higher-leverage—judgment, not janitorial labor. And the mistakes it catches look weirder, because they're the ones only a human would have found two years ago.
Vince Yeah, and this is the first time I’ve seen someone publish the actual CPU percentage saved by deleting a flag. That’s the kind of boring infrastructure metric I’ll frame one of these days.
Ava Oh, stop it.
Vince Come on, if we don’t get a boring-but-works tattoo soon, someone’s going to start a rumor. But for now, Ava, I want an agent that auto-cleans my unread Slack channels.