Amodei Put A Number On What Safety Has Cost Anthropic

Nathan Calvin of Encode AI published a critique of Anthropic's safety record, and Dario Amodei responded directly rather than through a comms channel — worth noting given how rarely frontier lab CEOs engage a named critic. His central argument is that Anthropic's safety positions have carried real commercial cost rather than being costless positioning, and the concrete example is a four-month delay to the Mythos model.

Four months is a substantial claim in a market where competitors ship on a rolling basis. Whether it persuades depends on something not disclosed: what the delay was for, and what would have shipped without it. A delay is evidence of a decision, not evidence the decision was correct, and Amodei is asking to be judged on a cost he can quantify rather than an outcome he can demonstrate. He also acknowledged that Anthropic mishandled some recent issues without enumerating them — a concession that makes the defence more credible than a flat denial, while leaving readers to substitute whichever controversy they already had in mind.

The timing sits inside a dense week. A draft IPO prospectus has been reported showing $16.3 billion in first-half revenue against roughly $9 billion in compute spending over six months, and more than $600 billion in future infrastructure commitments. The company has also been reported as potentially committing another $45 billion to Nscale, and has circulated a draft federal AI safety bill. A company that is simultaneously pre-IPO, committing hundreds of billions to infrastructure and drafting the legislation meant to constrain it attracts exactly this scrutiny — the safety record is not being examined in isolation, it is being examined as the thing that justifies the rest of the position.

Sources

The Neuron daily digest · Anthropic newsroom