Anthropic's latest update on its own engineering says something striking. Its people are now merging roughly eight times more code per day than they were in 2024.
Buried in the same paragraph is an admission. Lines of code is an imperfect measure that almost certainly overstates the real gain.
They reported the number anyway.
A metric that measures the wrong thing
I have never once seen more lines of code correlate with better software. Usually it's the opposite.
The best engineers I've worked with spend a good amount of their time removing code, not adding it. Deleting a redundant function. Collapsing three code paths into one.
Cutting a feature nobody used. That's the work that actually improves a system.
None of it shows up as a positive number on a lines-of-code chart. Most of it shows up as a negative one.
Anyone who has managed engineers knows lines of code is a weak productivity measure, and has known it for decades. Anthropic knows it too. They said as much, in the same update where they led with the number regardless.
Why report it if you know it's weak
The honest version of Anthropic's story would have worked fine on its own.
AI is now writing most of the code merged into a frontier lab's own codebase. Engineers are directing and reviewing rather than typing. Success rates on open-ended tasks have climbed sharply in six months.
That's an interesting claim on its own. It doesn't need a headline multiplier to carry it.
They used the multiplier anyway. Eight times more code is the kind of number that travels. It's simple, it's dramatic, and it fits a story people already want to believe.
Whether it's the right way to describe what actually improved is a separate question. Evidently not the one that decided whether to publish it.
What this should do to your scepticism
The company building the frontier model can't resist a vanity metric about its own use of it. That's not really a story about Anthropic.
It's about every other AI productivity number that crosses your desk this year. The employee sentiment surveys. The pilot programme readouts.
The vendor case study with the multiplier in the headline.
None of that means the underlying gains aren't real. Anthropic's engineers probably are getting more done. Uber's probably are too.
But "more done" and "more lines" aren't the same claim. The gap between them is where a number gets chosen because it travels well, not because it's right.
Read the caveat as carefully as the headline. Especially when the people reporting the number know better than anyone that it's weak, and reported it anyway.