Elon Musk said Anthropic is the current leader in AI and claimed the company "definitely" has an even stronger unreleased model ready to go.
The remarks came in a long interview with The Economist editor-in-chief Zanny Minton Beddoes, recorded on July 20 at Tesla’s Texas Gigafactory and aired on July 23. It was his first extended interview after SpaceX’s listing. The most concrete part of the conversation was not the ranking of models, but Musk’s AI governance proposal, which diverges clearly from the version put forward by Google DeepMind CEO Hassabis.
Musk’s case for Anthropic’s lead
Asked about progress by Chinese models, Musk said Moonshot AI’s Kimi K3 had come quite close to Fable 5, but that Fable 5 was still plainly the smartest model, adding that most people would agree with that assessment.
He then extended that view into a broader inference. If Anthropic’s Mythos was already ready in February, he said, the company must now have something stronger than that February version and could release it at any time. He stressed that this was not "almost certain" but "certain."
On the public record available so far, the premise behind that inference broadly lines up with known facts. Moonshot AI released Kimi K3 on July 16 with 2.8 trillion parameters, calling it the largest open-weight model to date, and published full weights on July 27. Moonshot’s own description acknowledged that K3 still trailed Fable 5 and GPT-5.6 Sol in overall performance, while beating both companies’ previous flagship models.
How Musk described OpenAI, Anthropic, and his own role
Musk also spent part of the interview tracing his relationship with the major AI companies. He said he originally co-founded OpenAI to counter what he viewed as Google’s near-monopoly in AI at the time. Anthropic later split from OpenAI, and in his telling the result was that "my actions actually accelerated AI" even though that was not his intention.
His comments on Anthropic CEO Dario Amodei were notably warm. Musk said Amodei is principled and genuinely cares about the world, and added that none of the Anthropic employees he had met had triggered his "bad person detector."
His description of OpenAI CEO Altman was much harsher. Musk repeated his core accusation that an organization that began as open source, nonprofit, and meant for the world ultimately became a closed-source for-profit company.
Musk’s governance proposal: let rivals test frontier models before release
This was the most specific policy section of the interview. Musk said he had spoken with Hassabis for several hours before Hassabis presented his own governance ideas, and that his suggestion had three layers.
- Leading AI companies should hold scheduled calls every one to two weeks to exchange information on safety and cybersecurity issues.
- Before releasing a new frontier model that is clearly stronger than existing ones, a company should let competitors test it through an API for one to two weeks. Musk emphasized that this means testing, not handing over the model itself.
- Only if a company is told about a risk and refuses to fix it should governments be notified to restrict release. In his view, the only governments with the capacity to act in practice are those of the United States and China.
The logic, as Musk framed it, rests on incentives. Government officials lack the first-hand technical judgment needed to decide whether a model should ship, he argued. Competitors, by contrast, can spot problems and have strong incentives to call them out, creating a form of mutual restraint.
He compared the idea to the Motion Picture Association rating system in the United States and said the mechanism should include Chinese frontier labs as well. On timing, he said six months is too long and that the process should begin immediately.
A lighter framework than Hassabis proposed
Set against the broader policy debate, Musk’s design differs sharply from the framework Hassabis outlined on July 14. Hassabis proposed a frontier AI standards body modeled on FINRA, funded by industry and overseen by the federal government. Under that structure, labs would submit models for review as much as 30 days before release. The system would begin as voluntary and later become a mandatory gate for deployment in the United States once it had proved effective. The body could also coordinate an industry-wide slowdown if needed.
Musk’s version is much lighter. It has no standing institution and no statutory force. At its core, it is a system of regular calls among peers plus a short pre-release review window in which rivals can probe a model through an API.
Asked whether the companies involved could realistically put aside their hostility, Musk said they could do so for the sake of the world.
The Mythos episode and who first raised the alarm
When the interviewer referred to the U.S. government using export controls to block Mythos, Musk added a detail: the cybersecurity risk was not first discovered by the government, but by Amazon, and Amazon was the one that called the White House.
That point broadly matches prior reporting, though the fuller sequence is more detailed. According to Axios and Fortune, Amazon researchers used a chain of prompts to get around safeguards and obtain cyberattack-related information that was supposed to be restricted. CEO Jassy then alerted senior White House officials, and at least five other companies called government agencies around the same time.
Anthropic released Fable 5 on June 9 and received a Commerce Department letter on June 12. The order barred all foreign nationals, including Anthropic’s own foreign employees, from accessing Fable 5 and Mythos 5. The company then disabled both models for users worldwide and did not restore access until July 1.
That left Amazon, which has invested more than $10 billion in Anthropic, in the unusual position of helping trigger a government block on its biggest AI portfolio company.
Some parts of the story remain disputed. Semafor reported that the U.S. government suspected groups linked to China had already used the vulnerability, but did not explain the basis for that claim. Anthropic said the White House had not mentioned any China-access issue in its communications with the company, and added that its products were already unavailable from within China.
Musk also faulted Amodei’s handling of the episode, saying he had effectively told the public the model was frightening and then released it anyway, a sequence that would naturally alarm people.
From extinction risk to a more fatalistic posture
One of the clearest shifts in the interview was Musk’s framing of AI risk. In the past, he signed an open letter calling for a pause in training and said AI could carry a 10% to 20% chance of causing human extinction. This time he said the risk is still not zero, but his stance sounded more resigned: if it cannot be stopped, the goal is to steer outcomes in the better direction. He even said that if there were a stop button, it might not be right to press it.
Beddoes pushed back with a direct question: if one of his rockets had a 20% chance of exploding, would he get on it? Musk said yes. His reasoning was that if the trajectory cannot be changed, the practical task is to reduce the odds of the worst outcome.
His broader predictions on AI, work, and money
Elsewhere, Musk stayed close to the style of forecasts he has made before. He said AI will exceed the combined intelligence of all humans in about five years, and that within 10 years the gap between humans and AI will resemble the gap between humans and chimpanzees.
In software engineering, he said AI already outperforms at least 90% of professional engineers and will eventually reach a "Stockfish level" gap, meaning the kind of distance where a phone can easily defeat the world chess champion.
From there, he argued that work will become optional. He used gardening as his analogy: you can buy better-looking vegetables, but growing your own is an act of personal meaning.
On money, he said bluntly that by 2036 money will no longer matter, and predicted that deflation rather than inflation will be the issue by then. If the output of goods and services grows faster than the money supply, he argued, handing out money directly would not push up prices.
An unresolved transition and a sharp turn into politics
This was also the weakest logical stretch of the conversation. Beddoes asked three times how a deeply polarized and fearful political reality today could transition into the world of abundance Musk was describing. His answer never moved much beyond saying the road would be bumpy.
The second half of the interview shifted toward European immigration, Starlink’s role in Ukraine, and arguments over DOGE cutting aid. The exchange was combative, but the two sides rarely met on substance.
The closing note was striking. Musk said he expects a British civil war in about 20 years, while the AI and robotics singularity will arrive within 10. By his own timeline, that leaves the political disputes in the second half of the interview looking less central than the AI trajectory he laid out in the first.