Anthropic urges coordinated global pause option for frontier AI

Anthropic’s “When AI builds itself” report and its call for a coordinated pause mechanism mark a turning point in how frontier labs talk about risk. Instead of vague references to alignment, the company published internal data showing Claude now authors the majority of its production code and that engineers ship roughly 8x more code per quarter than a few years ago—evidence that AI is already accelerating AI development. On that basis, the Anthropic Institute argues the world should build an evaluation-triggered “brake” that multiple labs could activate together if systems approach dangerous self‑improvement regimes. ([anthropic.com](https://www.anthropic.com/institute/recursive-self-improvement?utm_source=openai))

Strategically, this is both a safety intervention and a political move. A frontier lab publicly asking for the option to slow itself—and its rivals—creates pressure on peers like OpenAI and Google to articulate their own red lines. It also points away from unilateral moratoria (which Anthropic now says don’t work under competitive pressure) toward treaty‑like coordination schemes modeled on arms control. Critics will see regulatory capture in the making: only a handful of giants would sit at the “pause button,” and they have every incentive to freeze the frontier once they’re comfortably ahead.

For the AGI race, the net impact is ambiguous. If serious pause infrastructure emerges, it could slow the most extreme scaling runs. But the same analysis could be used to justify ever‑larger models in the near term, on the logic that we must “race to safety tools” before recursive self‑improvement arrives.