GPT-6 family models gained an internal field test on October 14, 2026, when OpenAI's developer blog published a retrospective from its Core Experimentation team. Using GPT-6 Astra and GPT-6.1 Sol with Codex to maintain a large codebase, five engineers working part-time over two weeks merged more than 1,400 cleanup pull requests and deleted over 800,000 lines of code. This was not a demo project; it was production code OpenAI itself runs every day.
Dead-code cleanup: keep the brief loose, demand the evidence
With GPT-5.6 Sol, engineers had to fence off the search area and spell out validation criteria before the model could start. With Astra, the team deliberately gave a single broad request to see whether it could decide where to look and what proof a removal needed. It reviewed past pull requests, internal discussion threads and project trackers, and checked production profiling tables to confirm the code was never executed, attaching a rationale to each deletion that sped up human review. The payoff was concrete: one service's container image shrank from 3.31GB to 1.72GB, a cut of about 48%, and roughly 400 visual snapshot variants were removed from the test suite. Working across long context while assembling its own evidence chain is the same long-task direction behind GPT-6.1 Sol raising its default context to 500k.
Stale experiments are harder: the risk is judging what can still move
Dead code can be rolled back; stale experiments are different, because their code may still run in production and eligibility checks can change at any time, which makes static analysis error-prone. The team let Astra prepare changes inside preset guardrails, each with evidence attached, and kept the final call with human reviewers. One example shows the judgment involved: a years-old internal feature flag that could override every other new-feature flag for testing mapped itself onto all features and caused significant CPU fan-out at runtime. The flag was marked permanent in the management console. Astra found the abnormal CPU cost but did not delete it on its own; it escalated the finding for human review, and after confirmation the removal saved an estimated 0.42% CPU in the main affected service. Across the effort, hundreds of stale experiments and feature flags were removed.
The cheaper tier's math: 92% of the findings at 14% of the cost
The team also ran a head-to-head test on a subdirectory of more than 6,000 files and one million lines of code, using the same prompt to audit stale experiments. Astra Ultra produced 142 findings at an API cost of $109.06 under standard pricing; GPT-6.1 Sol Ultra produced 131 findings for $15.12, roughly 92% of the findings at about 14% of the cost; Sol Medium produced 86 findings for just $5.27. Sol mainly missed constant flags that other flags depended on, a pattern that can be corrected in the prompt once identified. On the continuous integration front, about 30 optimization pull requests cut median CI completion time from roughly 10 minutes to 4.5 minutes, and one 63-test suite fell from 592.6 seconds to 29.8 seconds with every test retained.
For teams already using coding agents, the transferable part of this retrospective is not a magic prompt but the sequence: pilot on one small directory, require evidence with every proposed change, feed review feedback back into the prompt, and only then scale up. When choosing a model, measure the cheaper tier's hit rate first and reserve the most expensive model for the hardest audits. One caveat: every figure above is OpenAI's own internal reporting with no third-party replication, and gains on its codebase will likely shrink on a different one.
