Dario Amodei answers critic Nathan Calvin with nine safety calls Anthropic paid for

What Amodei posted

Anthropic chief executive Dario Amodei has answered one of the company's sharpest critics in public. In a long post on X, published at 02:26 UTC on 21 October 2026 (the evening of 20 October in California), he replied point by point to Nathan Calvin, a policy advocate at Encode AI, who had argued that Anthropic now behaves like a company whose first goal is winning the race to superintelligence.

Amodei opened by thanking Calvin and saying it "benefits the public to have people whose job is to scrutinize the actions of frontier AI companies." Then came the turn: "That said, I disagree with the thrust of your post." He set out nine decisions he said were "very costly to us commercially while having strong safety benefits," conceded three of Calvin's points and promised to look again at a fourth. He ended by saying he "will have more to say soon" about how Anthropic will "further live our values in this moment of rapid change."

The reply spread quickly. The Next Web reported it had been viewed more than 800,000 times, and The New Stack got a statement from Anthropic tying one of its key claims, a months-long delay to its Mythos-class models, to commitments in the company's Responsible Scaling Policy.

How the exchange started

The argument began with legislation. Anthropic has been shopping a draft federal AI safety bill that would let Washington halt a model only in narrow circumstances; The New Stack describes the threshold as a model judged to imminently risk killing or injuring more than 50 people or causing more than $1 billion in damage. Calvin, who worked on California's SB 1047 and SB 53, criticised the draft on 17 October, saying it seemed "far more preoccupied with protecting Anthropic from government overreach than it is with protecting the public."

Anthony Cimino, who leads Anthropic's federal affairs team, responded on 18 October. Calvin answered him the next day with a long post that, according to a timeline compiled by CellCog, went beyond the bill. "When Anthropic strays and makes mistakes, why do the mistakes systematically tend to guard and favor Anthropic's interests?" he asked, as quoted by The New Stack. He called his own list "deliberately one sided."

Calvin's nine items

As summarised by CellCog and The New Stack, Calvin's list had nine items. The first four were about scale: Anthropic's custom chip team for faster inference, more than $500 billion in computing deals as reported by The Information, Claude's fast-rising share of AI development work, and the large share of computing Anthropic spends on its own AI research.

The next two were about rules. Item five was version 3 of Anthropic's Responsible Scaling Policy, which in February dropped an earlier commitment not to train a system unless safety measures could be shown to be adequate in advance. Item six was Anthropic's separate frontier compliance framework, which Calvin said contains no binding rules on internal deployment. The last three were about how the company talks: a letter to Congress that was later walked back, the remark that you cannot do safety from second place, and the draft bill itself.

Amodei's nine costly decisions

Amodei's counter-list leads with politics. He said Anthropic backed SB 53 when AI regulation was unpopular and argued for a stronger version than passed. He pointed to biological safeguards strict enough that "biologists joke about how restrictive our classifiers are," and said rivals "sometimes sell against us by touting their more permissive bio capabilities." He said the company had "seen and blocked potential attempts to misuse our models for bioweapons."

The most concrete claim is about timing. "We first had Mythos-class models in February but didn't make them generally available until June," he wrote, offering them only to a small group of partners for cyber defence. "We likely sacrificed a significant chunk of a several-month model lead." Defenders did use that window: Bank of America says it added 1,000 contractors in the second quarter to patch flaws exposed by frontier models such as Mythos.

The rest of the list: his June essay calling for government power to block model releases; the company's stance during unspecified "events of February 2026," which he did not describe; support for chip export controls on China despite friction with partners; being, he said, the only frontier company to refuse sales to companies controlled by the People's Republic of China; long model cards and risk reports; and the decision this week to put 5% of the company into a public benefit fund. "Even if I conceded every item in your list, I would happily hold up mine next to it," he wrote.

What he conceded and what he defended

Amodei split Calvin's list into three groups. Items one to four, he said, "are all simply examples of what it means to compete at the frontier, and we stand by them." His argument is that the "race to the top" strategy, showing that AI can be built responsibly and pushing rivals to match, only works for a company at the frontier, so heavy spending on chips and computing follows from it.

He agreed Anthropic "could have handled your items 7, 8, and 9 better," putting them down to "rhetoric or policy communication that wasn't carefully written" in a company of more than 5,000 people. "I am of course the CEO, and I take responsibility." He called the scaling-policy change "the most substantive item," arguing the old rules forced effort on things that "in retrospect wasn't very important for safety" while missing others. On the compliance framework he said: "I promise to look into it further."

The pacing problem

One line may matter more than the lists. Discussing what to do "when the AI-driven acceleration gets fast enough that it is hard for safety to keep up," Amodei wrote: "We are nearing that point and that's what's driving some of our thinking about 'pacing.'" Anthropic's frontier safety roadmap, updated on 19 October alongside news that its automated red-teamer beat human bug hunters, sets 15 December 2026 for a proposal on how governments can support deeper access for embedded evaluators, and 15 January 2027 for publishing a policy framework on the pace of AI development or explaining why not, according to The New Stack.

That fits Anthropic's recent moves toward outside scrutiny, such as giving METR employee-like access to evaluate its systems. Amodei's promise of "more to say soon" suggests a further announcement on pacing is coming, though he gave no date.

What the reply does not settle

The exchange did not convince everyone. Sigal Samuel, who covers AI for Vox, asked Amodei for evidence that the strategy works, writing that "from the outside, it looks like Anthropic has done more on net to drive a race to the bottom," as reported by The Next Web. The race-to-the-top argument is hard to test, because there is no way to observe what rivals would have done without Anthropic.

Several of Amodei's claims are also hard to check. The bioweapon attempts, the commercial cost of the Mythos delay and the size of the business forgone in China are Anthropic's own accounts, and the draft federal bill at the centre of the dispute has still not been published in full. What the reply does establish is that Anthropic's chief executive now accepts responsibility for three communication failures, defends its spending as part of its safety strategy, and says the moment when progress outruns safety work is close.