Anthropic says Claude now writes 80% of its code, calls for an industry 'brake pedal'

Co-founder Jack Clark used a CNN and BBC Newsnight blitz to warn that recursive self-improvement is close — days after the lab filed confidentially for an IPO.

Anthropic now says Claude writes more than 80% of the code merged into its own systems, a share that sat in low single digits before Claude Code shipped in early 2025. The disclosure landed in a June 4 Anthropic Institute post titled “When AI Builds Itself,” co-authored by institute lead Marina Favaro and co-founder Jack Clark, which argues that recursive self-improvement “could come sooner than most institutions are prepared for.”

The supporting numbers are the actual story. Anthropic engineers now ship roughly 8x as much code per quarter as they did across 2021 to 2025. In April 2026, Claude shipped over 800 fixes that cut a class of API errors by a factor of one thousand; the supervising engineer estimated the work would’ve taken a human four years. GitHub data cited in the post puts industry-wide commits at about one billion across all of 2025 and 275 million a week by mid-2026, with the company’s COO saying GitHub is “pushing incredibly hard” on capacity. The 2026 run rate: roughly 14 billion commits.

Then came the media tour. On June 5, Clark told CNN’s Anderson Cooper that “I don’t have a brake pedal, and surely at some point in the future we might want that option,” reaching for a Cold War analogy: “We’ve done this before… they found ways to stabilize aspects of the nuclear arms race.” On BBC Newsnight he reframed it: “Right now, it’s like the AI industry has a gas pedal, but it doesn’t have a brake pedal.” Anthropic says it’ll convene governments, researchers and rival labs on a coordinated slowdown.

The timing is what critics can’t get past. Anthropic recently filed confidentially for an IPO that Scientific American reports could value it near $1 trillion. Bentley University’s Noah Giansiracusa was blunt: “I don’t think it’s a genuine call to slow down.” University College London’s Steven Murdoch told The Guardian the post “left out evidence that AI is close to self-improvement.”

It’s the 2008 TARP move applied to model labs: warn loudest about the systemic risk you’re best positioned to underwrite.

Anthropic says Claude now writes 80% of its code, calls for an industry 'brake pedal'

Sources