Anthropic publishes 'When AI builds itself', calls for coordinated pause option

Anthropic published an essay, “When AI builds itself,” arguing that automation of AI research and engineering is progressing fast enough to warrant preparing a coordination mechanism that would let frontier labs pause development jointly, while explicitly declining to claim that recursive self-improvement — AI systems capable of building their own successors — has arrived. “We are not there yet, and recursive self-improvement is not inevitable,” the essay states.

Its evidence was drawn from Anthropic’s own operations. The company reported that more than 80% of the code merged into its codebase is now written by Claude, up from “low single digits” before Claude Code launched in February 2025, with quarterly engineering output up eightfold between 2024 and mid-2026 — a figure Anthropic itself flagged as a measure of quantity rather than necessarily of quality. Separately, it charted the length of tasks models can complete unsupervised, from around four minutes for Claude Opus 3 in March 2024 to roughly 90 minutes for Claude Sonnet 3.7 a year later and about 12 hours for Claude Opus 4.6 by March 2026 — a trend it described as doubling roughly every four months, and projected forward to week-long task capability by 2027.

The proposed remedy was a verification system that would let “frontier AI developers verify that others globally have actually stopped or slowed” development, rather than relying on unilateral restraint. Anthropic acknowledged the mechanism’s central weakness itself: “training runs are far easier to conceal than missile silos,” and any credible pause would require multiple well-resourced labs in multiple countries agreeing to the same conditions simultaneously — a coordination problem with no working precedent in the industry.

The essay appeared days before Anthropic’s own confidential IPO filing and the launch of its Fable 5 and Mythos 5 models, prompting some observers to read it as risk-framing alongside a period of rapid commercial and capability expansion rather than a call to slow down in practice.