Anthropic says Claude now writes 80% of its code, urges global pause on AI development

Key Points

• Claude authored over 80% of Anthropic's production code in May 2026, up from single‑digit levels a year earlier.
• Engineers now merge eight times more code per quarter than in 2024; Mythos Preview boosts output about fourfold.
• Claude's success rate on complex engineering problems rose to 76% in May 2026, a 50‑point increase in six months.
• Automated Claude reviewer catches roughly one‑third of bugs that previously reached production.
• April 2026 research demo: nine Claude agents recovered 97% of a safety‑task performance gap in 800 hours of compute.
• METR data shows AI task horizons doubling every four months; Claude Opus 4.6 handles 12‑hour tasks.
• Claude Code accounts for 4.5% of public GitHub commits, generating about 2.6 million weekly.
• Anthropic Institute paper calls for a verifiable, multinational pause mechanism on frontier AI development.
• Three futures outlined: stalled progress, automated development with human direction, and full recursive self‑improvement.