Anthropic CEO Dario Amodei posted a long reply on X to researcher Nathan Calvin, who had accused the company of quietly loosening its own AI safety commitments. Amodei's core argument is the opposite: Anthropic has repeatedly given up business advantages to stay careful.
His headline example is the Mythos model family. The first of these models existed in February, yet broad access only opened around June, giving the company time to build cyberattack safeguards. Amodei concedes that the wait probably cost Anthropic a sizeable share of a lead over rivals that was measured in months.
What critics object to
Calvin points to huge compute contracts, Anthropic's growing reliance on AI to build its next models, and the rewrite of its Responsible Scaling Policy. The February version dropped a 2023 promise not to train a system unless safety measures could be shown in advance to be adequate.
Anthropic says the new policy focuses on transparency, public risk reports and independent evaluation, and Amodei calls the old rules too rigid. He still frames the strategy as a race to set the safety bar for the industry by staying at the frontier.
What it means for developers
The most capable models are now handed out in tiers. The verification program for security professionals has three access levels, with stricter vetting for broader use. Anthropic itself admits its filters sometimes wrongly block ordinary secure coding requests and offers a form to report them.
One more detail from the company's own reports: by its estimate, Claude wrote more than 80% of the code merged into Anthropic's main codebase as of May.
Our take
If you build websites or bots on Claude, two things matter here. New top models may arrive late and with restrictions, so don't plan a product around getting access on launch day. And if your service touches security topics, budget for false blocks and keep a fallback path. More in The New Stack.