Particle.news
Download on the App Store

Anthropic Says Claude Now Writes About 90% of Its Code

Practitioners describe a fast pivot to prompt-first workflows that recast hiring, triggering new safeguards for persistent quality errors.

Overview

  • Anthropic’s Claude Code and Opus 4.5 now generate roughly 90% of the company’s code, according to a spokesperson, with product lead Boris Cherny saying his own contributions have been 100% AI-written for more than two months.
  • Prominent engineers, including Andrej Karpathy, report a December 2025 “phase shift” where their work flipped from mostly manual coding to mostly agent-driven output, with some like OpenAI researcher “Roon” saying they no longer write code themselves.
  • Cherny detailed throughput gains, citing dozens of daily pull requests authored by Claude, and said the team even spun up a non-coder-facing tool, Cowork, in roughly a week and a half using the agent.
  • Hiring practices are shifting toward generalists as AI handles implementation details, and reporters note a decline in entry-level software openings as generative tools write more of the code.
  • Engineers caution that AI-generated code still makes subtle conceptual mistakes and leaves dead or overcomplicated code, prompting experiments with AI-on-AI review and continued human oversight, as Anthropic’s CEO projects broader adoption within six to twelve months.