Dario Amodei Answers Nathan Calvin: Anthropic's Nine-Point Defense, Checked

Dario Amodei has answered one of his most pointed critics in public. On October 21, 2026, the Anthropic CEO posted a long reply on X to Nathan Calvin, chief strategy officer at the AI policy nonprofit Encode. He called Calvin's work "constructive" and "honest," then wrote that he "disagree[s] with the thrust of your post." The reply had roughly 259,600 views when we captured it.

This post lays out what Amodei actually claimed, which parts we can check against the record, and which parts remain his own account. Calvin's original post, published October 20, is summarized below so the numbered items in Amodei's reply make sense.

TL;DR: the reply in one table

What did Nathan Calvin actually write?

Calvin's post went up on October 20, 2026, at 2:59 AM, and was addressed to Anthropic public policy staffer Anthony Cimino, who had answered Calvin's earlier criticism of the leaked federal bill (covered in our bill report). It had about 275,900 views when we captured it. Calvin wrote that the reply "is still not really engaging with the heart of my concern," then asked: what is Anthropic?

He offered three answers:

1. A very large tech company built to make employees and investors wealthy. He notes that an IPO at the reported 2 trillion dollar target would make it more valuable than Meta at 1.8 trillion.
2. An AI safety organization, as in the 2023 "Core views on AI safety" essay and Dario's recent call to "slow the pace at which we improve the capabilities of AI models."
3. An organization primarily trying to reach superhuman AI first, through systems that build their own successors.

His conclusion: Anthropic's recent actions look "most consistent" with the third view, with safety "fundamentally secondary." He lists nine actions, and these are the items Amodei's reply refers to by number:

Calvin also grants his list is "deliberately one sided" and credits Anthropic's published safety research and third-party collaborations. His sharpest question is the one critics keep repeating: "When Anthropic strays and makes mistakes, why do the mistakes systematically tend to guard and favor Anthropic's interests?" He ends with a quote from Anthropic co-founder Chris Olah about why outside critics matter, and says his remarks are not an attack on any individual.

Compare the two documents and the disagreement is plain. Calvin argues the pattern across all nine items reveals priorities. Amodei argues items 1 to 4 are not anomalies at all but the strategy, and concedes only the last three.

What is the "race to the top" defense?

Amodei's central argument is that much of what Calvin criticizes follows from a strategy Anthropic has held for years. "Our strategy has always been what I've called the 'race to the top,'" he wrote. Showing that the technology can be built responsibly raises the bar for everyone, and "the forces that drive the race to the top are hugely more effective if the company practicing them is at the frontier."

That is a coherent position, and it is the same one behind Anthropic's earlier safety framework. In February 2026 the company published Responsible Scaling Policy version 3.0, which states that the original theory of change, using capability thresholds to build consensus, "did not play out in practice" as hoped. So Amodei is not hiding that the plan changed. He is arguing that the goal stayed the same.

The weak point is that the argument is hard to falsify. If frontier competition is part of the safety strategy, then almost any frontier move can be placed inside it. Critics on X made that point within the hour. Reviewers should ask what evidence would count against the strategy, because the reply does not say.

The nine decisions Amodei says cost Anthropic money

Amodei wrote that his list of pro-safety, commercially costly actions in the last year "far outweigh[s]" Calvin's list. Here is each item, with what we can and cannot confirm.

Opposing groups and message bubbles around a gavel, illustrating the Anthropic policy debate over SB 53 and federal AI rules

1. SB 53 and other state laws. Amodei says Anthropic "played a key role" in California's SB 53 when that was politically risky, and that it argued for a stronger version than the one passed. Anthropic's support for the bill is public record. The claim that it wanted a stronger bill is Amodei's. For how this sits against the company's newer federal approach, see our report on Anthropic's pitched federal AI safety bill, where Calvin was quoted as "viscerally angry" and critics said the draft would preempt state laws like SB 53.

2. Strict biological safeguards. Amodei says Anthropic saw concerning results in bio uplift trials and blocked attempted misuse. He also concedes the cost: low false negatives mean many false positives, so that "biologists joke about how restrictive our classifiers are" and competitors sell against Anthropic on permissiveness. This matches the friction we covered in Fable 5 biology safeguards.

3. Holding back Mythos-class models. Amodei says Anthropic had Mythos-class models in February but did not make them generally available until June, offering them first to a limited set of partners focused on cyberdefense. He says this "likely" cost a chunk of a several-month lead. Our earlier coverage of Claude Mythos Preview and Project Glasswing documents the partner-first rollout, though not the internal February date.

4. Advocating for government power to block releases. He points to his June essay, Policy on the AI Exponential, which argued the government should be able to block model releases. Notably, the reported narrow, court-only version in the federal draft is a different design from a broad agency power.

5. February 2026. Amodei refers to "the events of February 2026" as an example where Anthropic's public stance "could have seriously damaged the company's future." He declines to re-ignite the conflict and does not name it in the post. We will not guess at the specifics.

6. Export controls on chips. He says Anthropic keeps advocating chip and equipment export controls despite friction with industry partners it depends on.

7. Restricting sales to PRC-controlled companies and fighting distillation. He claims Anthropic was "the first, and still only, frontier company" to restrict such sales, including via third countries. That "only" is a comparative claim we have not verified. On distillation, our report on Anthropic's Alibaba distillation case covers one concrete enforcement example.

8. Public reports and model cards. Long model cards, risk reports and advocacy, at a level of effort he calls implausible for a purely profit-maximizing firm. The August 2026 risk report is an example of the output.

9. The public benefit fund. Amodei says Anthropic announced this week it is putting 5 percent, and possibly more, of the company into a public benefit fund. See the fund's structure and the reaction to it.

What Amodei conceded

He did not claim a clean record. Three concessions stand out.

• General: "I also do agree that some of the examples you cite were mistakes by Anthropic."
• Items 7, 8 and 9: these "fall into the category of 'rhetoric or policy communication that wasn't carefully written or resulting from fast-changing circumstances not propagating from one team to another fast enough.'"
• Scale: Anthropic has over 5,000 people, and the problem "doesn't even necessarily represent a failure or mistake of any single person." Still, "I am of course the CEO, and I take responsibility."

Two readings of this are fair. One: it is an honest statement of how large organizations behave. The other: three separate communication failures in the same period suggest a pattern, and "it is a big company" does not say who owns fixing it. Which reading you choose depends on what you already believed.

The RSP rewrite: the most substantive dispute

Amodei called the RSP change "the most substantive item." His argument is that the old, rigid policy "created a lot of situations where we were technically required to do something that required a lot of effort but in retrospect wasn't very important for safety, as well as not covering some of the most important things."

What the new version emphasizes, per Anthropic's own announcement, is transparency and review: a published Frontier Safety Roadmap and Risk Reports "every 3-6 months," with third-party review "in certain circumstances." The roadmap's goals are described as "public goals that we will openly grade our progress towards," not hard commitments.

That distinction is the heart of the dispute. A hard commitment can be broken, which is visible. A graded public goal can be missed with an explanation. Amodei also points to the 2023 launch language, when Anthropic called the RSP an "early iteration" that would need "rapid iteration." That is true, and a fair rebuttal to anyone who says the company promised never to change it. It does not answer the narrower question of whether the change made commitments easier to meet.

On item 6, he said Anthropic's intent is to meet the letter and spirit of its regulatory obligations through its FCF while keeping the RSP for more experimental risk management. He added: "I promise to look into it further." Calvin's list names it the "frontier compliance framework" and says it contains no binding provisions on internal deployments; Amodei does not address that specific charge.

What critics said in reply

The replies under the post were mostly sceptical, and we present them as opinions, not facts.

• Michael Hinrichs argued that Calvin asked why mistakes always favor Anthropic and the "most you offer is 'big companies make mistakes.'" He called it "the longest non-answer I have ever seen."
• Dave Kasten (@David_Kasten) objected to calling a major policy rollout that would affect US law nationwide, and preempt the laws of Amodei's home state, merely a communication problem. His quote was truncated in our capture, so we are not stating its full claim.
• Others mocked the phrasing or disputed the word "democracies" in the export-control item.

Hinrichs's framing is the question to hold onto. Even if every item on Amodei's list is accurate, the list answers "has Anthropic done costly safety things?" and Calvin's question was "why do the errors point one way?" Those are different questions.

Where this fits in Anthropic's recent history

This exchange is one piece of a very busy month. Anthropic pitched a federal bill with independent evaluators, announced the fund, faced an NYC comptroller letter asking it to pause its IPO, and set out a plan for embedded third-party evaluators.

Amodei also hints at a coming shift. He writes that the company is "nearing" the point where AI-driven acceleration is fast enough that safety struggles to keep up, which is "driving some of our thinking about 'pacing.'" That links to his September essay We Must Pace the Frontier, where he wrote that "we must slow the pace at which we improve the capabilities of AI models," while adding that "pacing does not mean halting model training or technical progress."

He closes with a promise: "I will have more to say soon about how we are going to further live our values in this moment of rapid change."

What this means for what you build or pay

For most developers, the reply is background, not a product change. A few practical points do follow from what Amodei confirmed in his own words.

Crowd of opinion bubbles pointing up and down, representing public reaction to Anthropic's response to criticism

1. Expect continued bio friction. If your work touches biology, Amodei says the classifiers are tuned for very low false negatives and he intends to "keep doing it." Plan for refusals and use Anthropic's vetting route, which he names as the Life Sciences Verification Program.
2. Frontier access may arrive in tiers. The Mythos timeline he describes, partners first and general availability months later, is the pattern to plan around for the next top model.
3. China-linked customers face restrictions. If your company has PRC-controlled ownership, including through third countries, Anthropic says it restricts sales.
4. Regulation is moving toward evaluators. Between the federal draft and the pacing essay, third-party evaluation is the direction. Read our explainer on embedded evaluators to understand the term.

If you deploy agents and want a control layer that is independent of any lab's policy choices, AgentBeam, from the explainx.ai team, stops AI agents before they take dangerous actions.

Claimed versus verified

What to watch next

Three things will tell you whether this exchange changes anything. First, whether Anthropic publishes the "more to say" post Amodei promised, and whether it contains commitments or only values language. Second, whether the frontier-compliance-framework-versus-RSP question gets a public answer. Third, whether the federal bill text gets fixed on preemption, which the company's own federal affairs head has said was "not precise enough."

Criticism from groups like Encode is a normal part of how frontier labs are held to their stated strategy. Amodei said as much himself: "strong antibodies against corporate power are a healthy feature of a free society." Whether his reply answers the criticism is a call each reader can make from the primary text.

Related reading on explainx.ai

• Anthropic quietly pitches its own federal AI safety bill
• Dario Amodei: Policy on the AI Exponential
• Dario Amodei on pacing the frontier and embedded evaluators
• Anthropic's August 2026 risk report
• Anthropic's public benefit fund explained
• Fable 5 biology safeguards update
• Claude Mythos Preview and Project Glasswing
• Official: Responsible Scaling Policy v3.0

Details reflect Dario Amodei's post and our reporting as of October 21, 2026. Quotes come from the post as captured; engagement counts and replies change.