Anthropic Warns AI Self‑Improvement Is Accelerating, Calls for Global Pause

Anthropic released an unprecedented internal report that quantifies how its Claude model is accelerating AI‑driven software development. By May 2026, more than 80% of code merged into Anthropic’s repository originated from Claude, and engineer daily code output in Q2 2026 was eight times the 2024 level. The report frames this as the first concrete evidence of “Recursive Self‑Improvement” (RSI) moving from thought experiment to measurable reality.

External Benchmarks Confirm Rapid Progress

Public benchmarks indicate that AI model capabilities double roughly every four months, faster than the previous seven‑month cadence. For example, Claude Opus 3 (Mar 2024) completed a software task in four minutes that a human would need four minutes for; a year later Claude Sonnet 3.7 handled a task that would take a human one and a half hours, and Claude Opus 4.6 completed a twelve‑hour task in a single day. Similar acceleration appears in coding benchmarks (SWE‑bench) and research reproducibility tests (CORE‑Bench, METR), where scores moved from single‑digit to saturation within two years.

Internal Evidence of AI‑Powered Productivity

Anthropic’s internal data distinguishes two work categories: engineering (code, infrastructure, training supervision) and research (experiment design, result interpretation). Claude can take a loosely defined problem, propose a solution, and execute it, requiring only a high‑level goal from humans. As engineers gain experience, they shift from fixing specific bugs to defining broader objectives, and eventually to deciding quarterly development priorities.

Key metrics:

By May 2026, >80% of merged code was authored by Claude (previously single‑digit in early 2025).

Engineer daily code merges in Q2 2026 were eight times those in 2024.

In a March 2026 survey of 130 staff, median self‑reported productivity rose to roughly four‑fold when using Mythos Preview.

Claude‑written code quality is now comparable to human‑written code; by late 2025 the gap had largely closed.

Claude’s success rate on the most open‑ended tasks reached 76% in May 2026, a 50‑point increase over six months.

Claude can run experiments autonomously, improving code speed 3× (Opus 4, May 2025) to 52× (Mythos Preview, Apr 2026).

Automation also extends to code review: an internal Claude‑based reviewer catches about one‑third of potential vulnerabilities before they reach production.

Case Studies

In April 2026 Claude generated over 800 fixes for a specific API error, reducing the error rate by a factor of 1,000. Human effort to achieve the same would have taken four years. Another example shows Claude diagnosing a massive training‑job crash within two hours—a task that normally requires two to three days.

“About a year ago I devoted myself fully to the Claudifying project; it’s been a wild ride, and I haven’t written a line of code myself for roughly five months.” – Anthropic employee

Future Scenarios

Anthropic outlines three plausible futures:

Stagnation at a curve‑inflection point : Growth slows, and further gains require new architectures beyond Transformers.

Continued automation of AI development : AI handles most coding and experimentation, while humans set research direction. Organizations become dramatically more productive (a 100‑person team can accomplish the work of thousands).

Full recursive self‑improvement : AI designs and trains its successors autonomously, making compute the sole limiting factor.

Each scenario carries distinct risks. In the second and third cases, bottlenecks shift to human code review and decision‑making, echoing Amdahl’s law. The third scenario raises existential concerns about loss of human control, safety verification, and alignment.

Call for Global Coordination

Anthropic urges a coordinated, verifiable slowdown of frontier AI research. It proposes a “pause protocol” where leading labs across multiple nations agree to halt or slow development under mutually trusted verification mechanisms. The company plans a series of dialogues with policymakers, researchers, and civil‑society groups to explore such mechanisms.

Industry Context

Anthropic’s data reflect a broader industry trend. Other firms, such as the newly public “Recursive” AI startup (US$650 M funding) and DeepMind’s AlphaEvolve system, are also pursuing autonomous model improvement. Academic venues are acknowledging the topic: ICLR 2026 hosted a dedicated “Recursive Self‑Improvement” workshop.