Ask HN: I had Codex and GPT 5.6 Sol running for 12 days, 870k+ LOC. Now what?

An Ask HN poster reported running Codex together with GPT 5.6 Sol Ultra almost non-stop for about 12 to 13 days on a large extension to an existing SaaS product. The run burned through 502,122,866 tokens and produced more than 870,000 new lines of code. The author said earlier AI-assisted features in the 50,000 to 60,000 LOC range had already shipped and delivered real product value, but this batch was large enough that the practical next steps were no longer obvious.

At face value the run is a sustained multi-agent coding session: two models kept generating for roughly two weeks against one product surface, with total output more than an order of magnitude above the author’s prior successful AI features. Token volume near half a billion implies continuous context churn, repeated planning and revision passes, and a large generated tree rather than a single focused change set. The open question is less how the tokens were spent and more how much of the 870k+ LOC is reviewable, testable, and owned as production surface area.

For engineers and builders, the hard part is not generation capacity. It is integration: review load, regression risk, ownership of code nobody fully read, and the gap between “it compiles or demos” and “it is maintainable in a SaaS codebase.” Shipping 50–60k LOC of AI-written work is already a full team-scale review problem; 870k LOC forces a different workflow—layered ownership, automated checks, staged cutover, and explicit kill criteria for code that cannot be explained or tested.

The post sits in a familiar competitive frame on HN: tools such as Codex and GPT 5.6 Sol Ultra are already being used for multi-day, product-scale builds, not just autocomplete or small PRs. Peers who have shipped mid-five-figure AI features can map this post onto their own limits. Peers still treating agents as local helpers will read the token and LOC numbers as a stress test of process, not of model marketing.

Practical next steps from the numbers alone are operational, not visionary: freeze further generation; inventory modules by risk and coupling; require tests and human sign-off before any path is production-bound; measure how much of the 870k LOC is dead, duplicated, or unreferenced; and only then decide what to merge, rewrite, or discard. What to watch next is whether the author can turn a multi-week agent run into a bounded ship plan—or whether volume without review protocol simply becomes an expensive unmerged branch.