The Anthropic Institute · 2026-06-04 · major
Anthropic Institute's 'When AI Builds Itself' — Marina Favaro and Jack Clark Document 8× Engineer Output, 80% Claude-Authored Code, and Three Recursive Self-Improvement Scenarios Anthropic Says Could Land Before Society Is Ready
Anthropic Institute's first major essay argues compounding AI-on-AI development is already here — engineers ship 8x the code, Claude authored over 80% of merged production code in May, and full recursive self-improvement may arrive before verification regimes exist.
Anthropic argues AI is already automating its own development cycle, and full recursive self-improvement may arrive before any verification regime exists.
Key specs
| Engineer code throughput | 8x vs 2024 |
|---|---|
| Claude authored merged code (may 2026) | >80% |
| Claude error fix prs (april 2026) | 800+ |
| Open ended task success rate | 76% |
What is it?
An Anthropic Institute essay by Marina Favaro and Jack Clark — Anthropic's policy lead and co-founder — laying out the company's case that AI systems are increasingly designing, debugging, and shipping the next generation of AI. The piece sketches three near-future scenarios: a capability plateau with wide diffusion, compounding human-supervised efficiency gains, and full recursive self-improvement in which models design their own successors.
How does it work?
The authors back the argument with internal data: Anthropic engineers now merge roughly 8x as much code per quarter as in 2024, more than 80% of merged production code in May 2026 was authored by Claude, and Claude shipped over 800 PRs in April 2026 that cut a class of API errors by a factor of one thousand. They also cite a Claude Code session study where models matched or beat human researchers on next-step suggestions during open-ended AI safety research. The essay argues detection of unilateral defection on a coordinated pause is harder than for nuclear arms — training runs hide easier than missile silos — and proposes building verification tooling now while the window is open.
Why does it matter?
It's the first time a frontier lab has put a concrete near-term timeline on recursive self-improvement and published internal productivity numbers to back it. Jack Clark separately told Axios there's a >50% chance an AI system can autonomously build a better version of itself by end of 2028. The piece reframes the AI-safety conversation around timelines policymakers and societies have not begun preparing for.
Who is it for?
AI policy researchers, lab leadership, governance funders, and engineers tracking labor-displacement curves
Try it
https://www.anthropic.com/institute/recursive-self-improvement