Yesterday in AI: 05 June 2026 — Claude Now Writes 80% of Its Own Code
Anthropic's 'When AI Builds Itself' report reveals Claude authors 80%+ of production code; Apple approves first AI agent inside iMessage; Congress tables 269-page federal AI act with 3-year state preemption.
By OMC Editorial on 2026-06-06
TL;DR — Anthropic's "When AI Builds Itself" report reveals Claude authors 80%+ of its production code and engineers ship 8× more per quarter, with the company calling for a global AI pause mechanism; Apple approved Poke as the first standalone AI agent inside Messages for Business, reaching a billion iMessage users days before WWDC; Congress unveiled the 269-page Great American AI Act, proposing three years of preemption on state AI development laws.
---
1️⃣ Claude Writes 80% of Anthropic's Own Code — Now Asking for a Global Pause Button
- What: Anthropic published "When AI Builds Itself," a report revealing Claude now authors more than 80% of the company's merged production code — up from single digits before Claude Code launched in February 2025
- Why it matters: Anthropic is the first major frontier lab to quantify recursive AI development at this scale and to publicly call for a verifiable, multilateral pause mechanism before AI systems can design their own successors
- Key number: 80% of production code authored by Claude; engineers now ship 8× more code per quarter than any point between 2021 and 2024
The report, authored by co-founder Jack Clark and Anthropic Institute director Marina Favaro, draws on public benchmarks and previously unreported internal data. Claude's success rate on open-ended coding tasks reached 76% in May — up 50 percentage points in just six months. The report distinguishes between "AI-accelerated development" and true recursive self-improvement, warning the threshold could arrive faster than governments are prepared for.
Anthropic stops short of claiming recursive self-improvement has begun. Instead it argues the inflection point is close enough to warrant pre-emptive infrastructure: a coordinated, verifiable mechanism all major labs could jointly trigger — analogous to existing treaty frameworks for nuclear and biological weapons.
On the same day, a multi-hour outage took down claude.ai, Claude API, Claude Code, and Claude Cowo