Anthropic says its engineers now merge 8x more code a day. The thread did the math

The number Anthropic chose to lead with

Anthropic's essay on recursive self-improvement argues that for most of AI's history humans drove every step of development, but the company is now "delegating a growing share of AI development to AI systems themselves." The headline evidence: in the second quarter of 2026, Anthropic engineers merged 8x as much code per day as they did in 2024, with Claude authoring more than 80% of merged production code.

The piece stacks up other curves too: task-completion horizons doubling every four months (Claude went from 4-minute tasks to 12-hour tasks in a year), SWE-bench going from single digits to saturation in two years, and internal optimization experiments accelerating from ~3x speedups in May 2025 to ~52x in April 2026. It lays out three futures and calls the middle one — compounding efficiency gains with humans still steering — "likely."

The footnote Anthropic wrote against itself

Buried in the same essay is the caveat that the thread seized on: "Lines of code is an imperfect measure… So 8x lines of code/engineer/day in the second quarter of 2026 is almost certainly an overstatement of the true productivity gain." Employee surveys, the essay admits, reported only a ~4x subjective bump.

The HN crowd — many of them senior reviewers — went further. A developer who declined to review a 40,000-line AI pull request described the genre:

"One of my co-workers just asked me to review his pull request… 600 files were touched, over 40k lines of code added… what engineer could write 40k lines of code in a week?" — an engineer who refused to review it

"I just watched copilot today turn a 8 line fix into 500 lines, so, yeah, verbosity is a big side effect" — a developer in the thread

The sharpest cut came from someone pointing out the metric runs in both directions:

"If you believe negative lines of code is the goal, then they've gotten 8x worse!" — a skeptic doing the inverse math

Max Woolf, a senior data scientist at BuzzFeed, offered the constructive counter-pattern: don't measure lines, measure whether the model can iteratively optimize real benchmarks "without cheating or causing regressions." And a frequently-cited rule of thumb resurfaced:

"AFAIK, the only correlation with LoC that's got solid evidence is this: the number of bugs correlates with LoC." — a commenter citing the research

The timing nobody missed

The top comment didn't argue the metrics at all — it argued the calendar. The most-upvoted reply read the essay as part of an IPO roadshow: if AI is solving problems of "great magnitude," why hasn't Anthropic pointed it at cancer or fusion instead of at its own code review queue? Others found the moral framing worse than the marketing — the "we must build the dangerous thing before someone else does" argument struck one reader as "a coward's argument."

Read the curve, then read the reviewer

Anthropic's curves are real and the task-horizon doubling is genuinely striking. But the company picked the one headline number it then spent a footnote disowning — and the people who actually merge the code read 8x more lines as 8x more to review, not 8x more shipped. The most honest line in the whole exchange was Anthropic's own: this is "almost certainly an overstatement." When the vendor and its harshest users agree the metric is inflated, the burden of proof sits with the curve that isn't about volume.