Anthropic calls for a pause (sort of)

In April, an Anthropic engineer pointed Claude at a backlog of bugs nobody else wanted to touch and walked away. The model came back with more than eight hundred fixes and a class of API errors cut by a factor of a thousand, work the engineer reckoned would have taken a human four years. He was not bragging. He was describing his own obsolescence, politely, in the past tense.

The anecdote sits inside a document the Anthropic Institute published this week, a polished, unnerving, faintly triumphant argument that artificial intelligence has begun to build artificial intelligence. Its authors are Marina Favaro and Jack Clark, the latter a co-founder of Anthropic and its head of policy, which is the first tell. A safety paper, it turns out, is being written by the man whose other job is keeping Washington calm.

The figures are the sort Anthropic does not usually let out of the building. More than 80% of the code it now merges is written by Claude, up from low single digits before its coding tool shipped in February 2025. On an internal test that has the model rewrite training code to run faster, the best system went from roughly tripling the starting speed last spring to a fifty-two-fold gain by April. The typical engineer now ships eight times the code he did in 2024 — a figure Anthropic itself calls flattering, since lines of code count volume and say nothing about worth.