Anthropic says Claude now writes over 90% of its code and wants the world to have an AI pause button

The headline figure in circulation is 80 percent, but Anthropic's own disclosure puts it above 90 percent, a distinction that matters because it shifts the framing from 'AI assists engineers' to 'engineers assist AI.' The pause mechanism proposal is also not a unilateral commitment but a conditional one, requiring competing labs to match it before Anthropic would comply.

This lands four days after Anthropic's confidential S-1 filing, covered here from multiple angles including The Verge and WIRED in early June. Those pieces flagged the core tension: public markets reward near-term revenue growth, while Anthropic's brand equity rests on safety credibility. Proposing a global pause button while disclosing that your model already writes most of its own code is a precise response to that tension. It lets Anthropic tell public market investors it is scaling fast while telling regulators and safety-focused stakeholders it is not scaling recklessly. Whether that framing holds under SEC disclosure scrutiny is a different question.

Watch whether any named competitor, OpenAI, Google DeepMind, or xAI, publicly responds to the pause proposal within the next 60 days. Silence or rejection would expose the proposal as a positioning move rather than a genuine coordination attempt, which would matter considerably once Anthropic's S-1 becomes public.