Anthropic reports AI now writes over 80% of its merged code. Engineers there ship 8x more code per quarter than in 2021-2025. AI success on open-ended coding tasks reached 76% in May 2026. Claude achieved a 52x speedup in model training optimization tasks. AI can now run end-to-end research projects with minimal human input. Anthropic warns recursive self-improvement may arrive sooner than institutions expect. It could bring major benefits but also heighten control risks. The trend is not yet inevitable. Public benchmarks show AI task duration capability doubling every four months. Internal data shows AI handles most engineering and some research tasks autonomously.
What changed
Humans wrote nearly all code and directed every research experiment at Anthropic.
What it unlocks
AI systems can now autonomously write, test, and optimize large portions of their own development workflow.
- 8x code per engineer per quarter
- 80% of code authored by Claude
- 76% success rate on open-ended tasks
- 52x speedup in model training optimization
Sources