Anthropic CEO Dario Responds in a Long Post to Doubts: Mythos-Class Model Kept Under Wraps for 4 Months
Anthropic CEO Dario published a long post on X responding to doubts raised by AI safety investigator Nathan Calvin, admitting for the first time that the company had already trained a high-end Mythos-class model in February this year, but kept it under wraps for about 4 months because its cybersecurity defense mechanisms were not yet fully developed, only opening it to the public in June. He also listed the costs the company has proactively borne for safety, including normal research being mistakenly blocked due to biological risk filters, pushing for California's SB 53 legislation, and plans to contribute 5% of the company's shares to a public interest fund.
Body · AI translation
In the small hours of Silicon Valley, the air was thick with the smell of gunpowder.
Just moments earlier, Anthropic chief Dario had, in a rare moment, lost his composure on X.
He personally threw out a nearly thousand-word hard-hitting essay, directly pushing back against the most cutting question the outside world has raised about Anthropic—
In response, Dario for the first time revealed—
A fully trained Mythos-class advanced model was kept under wraps for about four months before the public was allowed to touch it.
He laid out, item by item, the money his company had "deliberately lost over the past year for the sake of safety."
In an industry where a model can rewrite the rankings every few weeks, four months is enough time for round after round of new models to take the stage.
Choosing to wait very likely cost Anthropic a considerable portion of what had originally been a months-long lead.
"Anthropic just wants to win the superintelligence race"
This time, the person who forced Dario to show his true colors was Nathan Calvin, a well-known investigator in the AI safety field.
A few hours earlier, Calvin's thousand-word polemic had instantly set the entire AI world ablaze.
He pulled out a long list, dissecting with surgical precision the "nine sins" of Anthropic over the past year.
A massive $500 billion compute order, a secretly assembled in-house chip team, letting Claude take over 26% of internal R&D tasks for "self-iteration," quietly revising safety commitments...
Calvin's subtext was razor-sharp: drop the act—you just want to build superintelligence before everyone else!
Calvin and several policy researchers were furious after reading the draft, believing that the bill Anthropic was pushing behind the scenes looked like a call for regulation but actually concealed a hidden agenda—
Anthony Cimino, Anthropic's head of federal affairs, had wanted to post something over the weekend to put out the fire, but instead completely lit Calvin's fuse.
Calvin fired off a series of soul-searching questions: What kind of company is Anthropic, really?
If you go by what executives say in their open letters, it is a "safety organization" always ready to hit the brakes for humanity.
Not long ago, Dario had publicly warned that within the next 6 to 12 months, AI clusters could take over the entire internet, and called on the whole industry to slow down.
Calvin delivered a cold jab: if the nine things above had been done by OpenAI, Anthropic's employees would probably have jumped out long ago to curse them out and treat it as ironclad proof that their rival was "blinded by greed and indifferent to humanity's fate."
Dario lays out the books late at night
How much revenue have we cut ourselves off from for safety?
Faced with accusations that threatened to overturn the company's moral foundation, Dario did not hide.
At the start of his response, Dario first showed some high emotional intelligence, thanking Calvin for his oversight, but once the pleasantries were over he went straight for the knife.
Right after that, he laid out in one breath nine "self-harming moves" that would give countless Silicon Valley investors chest pains.
A god-tier model in hand, deliberately kept under wraps for 4 months
This is the most information-dense bombshell in the entire essay.
Dario publicly admitted for the first time that as early as February this year, Anthropic had already trained a "Mythos-level" advanced model.
But Anthropic held back.
At the time, the cybersecurity defense mechanisms for models of that level had not been fully polished, so they chose to open it only to a very small number of partners, specifically to strengthen cyber defense testing.
It was not until June this year, when the safety guardrails were in place, that the model was officially opened to the public.
A full four-month lead window, in a race where every second counts, was cut off by their own hands.
Dario said plainly that it was precisely because of this four-month buffer that many enterprises were able to shore up their security infrastructure before AI models with powerful cyberattack capabilities became widespread.
On June 9 this year, Fable 5, which Anthropic launched to ordinary users, belongs to this tier and comes with strict protections.
Mythos 5, released at the same time, first gave a small number of cyber defense and infrastructure partners partially unrestricted capabilities.
So during the "shelved" period there was still restricted use; when it opened in June, there were also different access boundaries.
Of course, he did not provide a quantifiable accounting of exactly how much revenue the wait cost or how much risk it avoided.
"Zero tolerance" for biological risk
The cost in the biological domain was especially painful for Dario. He himself has a biology background.
He said the company had seen worrying results in biological capability tests and had blocked potential attempts at biological weapons misuse, and therefore applies strict protections to high-risk uses.
The problem is that filters have a hard time accurately reading each person's intent.
Letting a dangerous request slip through is a false negative; blocking legitimate research is a false positive. To minimize the former, Anthropic accepted a large number of the latter.
Legitimate researchers were frustrated, and competitors would use "fewer restrictions" as a selling point.
Users bear this cost too.
Anthropic's previous official explanation confirmed that when Fable 5 triggers the biosafety filter, it hands the request to a less capable model.
An August update reduced biology-related fallbacks by about 85% in company testing, but still retains restrictions on dual-use research.
Another item on the ledger is regulation.
Dario said Anthropic pushed for legislation such as California SB 53 and even supported versions stricter than the one that ultimately passed. In June this year, he also publicly argued that the government should have the power to block model releases.
That amounts to supporting an external force holding the switch to its own products.
Finally, the 5% equity stake. Just a few days ago, Anthropic announced plans to put 5% of the company's shares into a public interest fund, pending shareholder approval. Once the company's valuation exceeds a specified benchmark, the board may also add 2% at its discretion for every 50% increase.
Why does a company that insists on safety still desperately buy compute and accelerate R&D?
The core logic Dario gives is called "race to the top"—competing toward higher standards.
Calvin's worry is also direct: if every round of competition must be won, how much more can safety make the company give up?
This article comes from the WeChat public account "Xin Zhi Yuan," author: ASI Revelations, editor: Taozi, published by 36Kr with authorization.