Anthropic CEO Dario Amodei has answered a critic on X with a nine-item list of actions he says made the company safer and cost it money. The reply, a thread Amodei posted on 21 October 2026, opens by thanking Nathan Calvin, whose earlier post criticized the company’s conduct. Everything below is Amodei’s own characterization of events, reported by Dario Amodei, on X. The thread contains no outside verification and no dollar figures.
The reply lands in a busy week for Anthropic’s safety reputation. METR’s review of the company’s risk report and an essay critical of the labs have kept the question in front of readers, and Amodei’s list reads as an attempt to set the terms of that argument himself.
Amodei says the items covered “the last year,” and that they came “while having strong safety benefits.” In his telling, the nine are:
- Helping shape SB53 and other state laws when, he says, regulation was unpopular, and arguing for a stronger SB53 than the one that passed.
- Building strict biological safeguards. He says the filters also block legitimate work, to the point that biologists joke about them and competitors pitch against Anthropic on it.
- Holding back its Mythos-class models. He says Anthropic had them in February, kept them with a small group of partners working on cyberdefense, and opened general access in June after adding stronger cyber safeguards. He estimates the delay probably gave away a large part of a lead that lasted several months.
- Arguing in his June essay “Policy on the AI Exponential” that government should be able to block model releases. He says the company was attacked intensely for it.
- Taking a public stance in February 2026 that, he says, put the company’s future at real risk. He declines to elaborate.
- Pushing for limits on exporting chips and chip-making tools to China, despite friction with industry partners it depends on.
- Restricting sales of AI services to companies controlled from the PRC, including through third countries, and cracking down on distillation. As far as he knows, he writes, Anthropic was the first and is still the only frontier company to do the former. That is his claim, not a verified ranking.
- Publishing reports, announcements, and long model cards that he says no pure profit-maximizer would bother to produce.
- Announcing this week that 5 percent of the company, and possibly more later, will go into a public benefit fund.
Amodei does not argue that these choices were free. He concedes that restricting access had a broader cost, since the public might have gained from faster progress in biology. He also anticipates the objection that Anthropic should not decide who gets access at all, and says he agrees in principle: government ought to make those calls, which is why the company lobbies for regulation.
On Calvin’s own list, he concedes ground in one place and holds it in others. He calls three items “rhetoric or policy communication” that was not carefully written, says he takes responsibility as CEO, and notes Anthropic has more than 5,000 employees. The change to the Responsible Scaling Policy he treats as the most substantive point. In his view the old, rigid version forced effort that did not matter much for safety, while the current one brought risk reports, roadmaps, and independent external review. On the company’s separate framework for meeting regulatory obligations, he promises to look further into the concern.
The structural weakness is that the scorecard has one author. Amodei picks the period, picks the items, and judges the cost to his own company, while giving no revenue figure for any of them. He also says the comparison is partly subjective and that his list mostly concerns misuse rather than loss of control, which he allows some readers weigh more heavily.
For companies buying Claude, the biology filters are the concrete part: Amodei confirms in writing that they over-block, and he says the pain falls on biomedical users. Anyone in that field should test the current safeguards on real workloads, and hold Anthropic to the “more to say soon” he promised on how it will live its values.
Source: Dario Amodei, on X, in a reply to Nathan Calvin, posted on 21 October 2026.
