Anthropic urges global 'temporary pause' as Claude nears recursive self-improvement

Anthropic published a report and blog post detailing what it calls accelerating progress toward recursive self-improvement, warning that the pace of capability gains is outstripping the industry's ability to keep humans in the loop. The headline claim: as of May 2026, more than 80% of code merged into Anthropic's production codebase was written by Claude, a dramatic jump from early 2025. Internal experiments cited Claude-powered agents completing AI-safety research autonomously and a Mythos Preview model achieving a ~52x speedup on code-optimization tasks alongside 76% success on hard engineering problems.

Mechanically, the argument is that agentic models increasingly improve the very tools and pipelines used to train and evaluate their successors, creating a feedback loop. Anthropic framed this as cause for a coordinated 'temporary pause' so policymakers and labs can agree on guardrails before capability curves steepen further. The company says it wants to host discussions rather than impose a unilateral halt.

Competitively, the move lands as rivals push the opposite direction—Microsoft touting in-house superintelligence ambitions and OpenAI racing toward an agent 'super app.' Critics, including voices cited by Scientific American, argue the 80%-of-code framing is as much marketing as warning, noting that code volume is a poor proxy for productivity (a point Keras creator François Chollet made bluntly on X). Skeptics also see a pause call from a frontier lab as a way to slow competitors while preserving its own lead. Readers should watch whether any other major lab endorses the pause, or whether it remains an Anthropic-only position.