Anthropic Warns: AI Self‑Improvement Is Accelerating Faster Than Expected – Calls for a Global Pause

Anthropic announced a major internal finding: the Claude model is accelerating AI development, creating a pathway toward recursive self‑improvement where AI can design successors that surpass its own capabilities. The company’s internal report, based on engineering metrics, employee surveys, and experiments, shows that by May 2026 more than 80% of code merged into Anthropic’s repository was authored by Claude, and engineer daily code output in Q2 2026 was eight times that of 2024.

The report emphasizes that this is the first concrete evidence bringing the long‑standing thought experiment of recursive self‑improvement into real‑world discussion. Anthropic also publicly calls on global peers to pause large‑model research.

External benchmarks support the trend: AI model task‑completion times are halving roughly every four months, with Claude Opus 3 (Mar 2024) completing a four‑minute software task in minutes, Claude Sonnet 3.7 (Mar 2025) handling a one‑and‑half‑hour task, and Claude Opus 4.6 (Mar 2026) tackling twelve‑hour tasks. Similar acceleration appears in coding benchmarks (SWE‑bench) and research reproducibility tests (CORE‑Bench, METR).

Anthropic’s internal data further detail the workflow shift. Engineers now receive high‑level goals and Claude generates detailed solutions, reducing human effort to guidance and review. Code quality has improved, with Claude‑written code passing automated reviews and fixing bugs that would have taken humans years to resolve. In a May 2026 survey of 130 staff, perceived productivity rose roughly four‑fold when using Claude.

Performance metrics show Claude’s code‑generation speed increasing dramatically: Claude Opus 4 (May 2025) was three times faster than its predecessor, and Claude Mythos Preview (Apr 2026) achieved a 52‑fold speedup over baseline. Success rates on open‑ended tasks rose from 26% to 76% within six months.

The authors outline three future scenarios: (1) the acceleration curve plateaus, possibly due to diminishing returns or architectural limits; (2) AI‑driven automation continues, making organizations dramatically more productive while human roles focus on direction and evaluation; (3) full recursive self‑improvement emerges, allowing AI systems to autonomously design and train successors, making compute the sole driver of progress.

They discuss bottlenecks such as human code review (an Amdahl’s‑law effect) and the need for coordination mechanisms to verify pauses, noting that verification is harder for AI than for other technologies. Anthropic proposes a verifiable global slowdown protocol and plans dialogues with policymakers, researchers, and civil society.

Industry context shows the trend is not isolated: other firms (e.g., Recursive) are pursuing self‑improving AI, academic workshops on recursive self‑improvement appear at ICLR 2026, and DeepMind’s AlphaEvolve demonstrates similar evolutionary programming capabilities.

In conclusion, Anthropic presents three plausible outcomes without committing to optimism or pessimism, urging the community to consider coordinated slowdown while acknowledging the uncertainty of controlling increasingly autonomous AI systems.