Research

Anthropic Warns of Recursive Self-Improvement as Claude Writes Over 80% of Its Code

On June 4, 2026, Anthropic’s Institute published “When AI builds itself,” documenting how AI already accelerates AI development and arguing that full recursive self-improvement could arrive sooner than institutions are prepared for. Internal metrics include more than 80% of merged Anthropic code authored by Claude as of May 2026, roughly 8× code merged per engineer versus 2024, and open-ended task success rising to 76%. The lab calls for research into verifiable global slowdown or pause options without a unilateral freeze that merely hands the lead to less cautious actors.

Tech Insights Reporter Jun 4, 2026 7 min read San Francisco, CA

TLDR

Anthropic Institute’s June 4, 2026 essay “When AI builds itself” presents public and internal evidence that AI systems are already speeding AI R&D—and that a future of full recursive self-improvement (AI designing its own successors) may arrive faster than policy can adapt. Concrete Anthropic metrics include Claude authoring over 80% of merged production code by May 2026, engineers shipping about 8× more code per day than in 2024, and sharply rising success on open-ended tasks. The piece sketches three futures and argues for building verification systems that would make a coordinated, multi-lab pause possible, not a solo freeze.

Evidence Anthropic highlights

External trends

Long-horizon task reliability doubling roughly every four months (METR-style horizons), from minutes (Opus 3, 2024) to multi-hour and 12-hour tasks (Opus 4.6 era).
Rapid saturation of coding and research reproduction benchmarks (SWE-bench, CORE-Bench); Mythos Preview described as working “at least” 16 hours on METR’s hard end.

Internal Anthropic data

>80% of lines merged to Anthropic’s codebase authored by Claude as of May 2026 (up from low single digits before Claude Code).
Typical engineer merging ~8× as much code per day in Q2 2026 vs 2024 (lines-of-code caveat acknowledged).
March 2026 poll: median research-team estimate ~4× output with Mythos Preview vs no AI (with self-report caveats).
Open-ended task session success 76% in May 2026 (+50 pp in six months).
Experiment-optimization loop: ~3× speedup (Opus 4, May 2025) → ~52× (Mythos Preview, April 2026) on a fixed mini training-code task.
Automated Claude code review would have caught roughly one-third of past claude.ai production-incident bugs retrospectively.
Research judgment probe: models beating human “next step” choices rose from 51% (Opus 4.5, Nov 2025) to 64% (Mythos Preview, April 2026) on selected detour moments.

Policy stance

Anthropic argues a coordinated, verifiable slowdown could be desirable if peers at the frontier can prove they stopped—otherwise a unilateral pause mainly helps less cautious competitors. The Institute says it will research verification/coordination mechanisms and convene policymakers, researchers, civil society, and other labs. The essay’s timing—days after Anthropic’s confidential S-1 filing—drew market commentary, but the primary piece is framed as capability evidence and governance research, not an IPO prospectus.

Why this story matters

This is rare, quantitative disclosure from inside a frontier lab on how deeply AI already automates its own engineering. If >80% of a leading lab’s code is model-written and experimental loops are superhuman, the bottleneck shifts to human judgment, review, and verification—and eventually to whether models can set research directions. That is the practical bridge from today’s coding agents to the recursive self-improvement scenarios governments and markets are only beginning to price.

Sources

Anthropic Institute: “When AI builds itself” (anthropic.com/institute/recursive-self-improvement, published June 4, 2026). Primary essay with metrics, scenarios, and pause-option framing.
Reuters and other same-day coverage of Anthropic’s coordinated-pause discussion (June 4–5, 2026).
Related Anthropic research links cited in the essay (METR horizons, Glasswing, alignment work).

Featured Image Alt Text

Abstract recursive loop of AI agents writing code and training models, with Anthropic Institute branding.