Anthropic Warns: AI Is Self‑Evolving—Should the Industry Pause?

Anthropic internal data on Claude

By May 2025, >80 % of Anthropic’s codebase was authored by Claude, up from single‑digit percentages before Claude Code release. Engineer‑level code merges per quarter are eight‑fold higher than the 2021‑2025 average.

Claude’s success rate on open‑ended programming tasks rose from 26 % six months ago to 76 % today (a 50‑percentage‑point increase). Engineers report code quality now matches human engineers and expect it to surpass human performance within the year.

AI‑independent task duration metric

Anthropic defines “AI‑independent task duration” as the time a human would need to complete a software task. Reported values:

Mar 2024 – Claude Opus 3 completed a task in ~4 minutes (human baseline).

Mar 2025 – Claude Sonnet 3.7 required 1.5 hours.

Mar 2026 – Claude Opus 4.6 required 12 hours.

Mythos prototype can sustain at least 6 hours of continuous work, reaching the upper limit of the METR testing framework.

The doubling period for this metric shortened from 7 months to 4 months, implying tasks lasting several weeks could be handled by 2027.

Claude as automated code reviewer

All code changes are routed through Claude for automated review (bug, security, and defect detection). Retrospective analysis indicates that about one‑third of bugs that caused incidents on Claude.ai would have been caught by this step.

Research‑level acceleration

Task: evaluate whether a weaker model can reliably supervise a stronger one.

Human baseline: two researchers reduced the performance gap by 23 % in one week.

Claude baseline: after ~800 hours of compute costing ≈ $18 k, Claude reduced the gap by 97 %.

Future scenarios for recursive self‑improvement

Stagnation. Exponential curves could flatten into an S‑curve; breakthroughs in judgment or new architectures may be required, or physical limits (energy, chips, compute) could dominate.

Continued acceleration with human steering. Organizational efficiency scales exponentially, but bottlenecks shift to code review and integration, analogous to Amdahl’s law at the company level.

Full recursive self‑improvement. Development speed becomes limited only by compute; humans remain in supervision and verification roles; capability could spill over to other scientific domains.

Call for verifiable slowdown mechanism

Anthropic states it would slow down or pause development if a verifiable mechanism guaranteeing that no lab is secretly racing ahead were established.

Comparison with OpenAI

OpenAI’s recent blog notes early signs of AI‑driven self‑acceleration, heightened competition, and governance challenges.

Reference: https://www.anthropic.com/institute/recursive-self-improvement