Transcription
Mythos, Fable, Opus 4.8. Anthropic wants everyone to know that they're steering us into a brave new world. They've now officially taken over OpenAI on almost every front, becoming the most powerful AI company in the world.
"I was watching this graph for a while and I said, 'Oh yeah, we'll probably become the, you know, the the AI company with, you know, the the most revenue and the most valuation sometime around this time.' And and indeed indeed it has happened."
Mythos is marketed as a quantum leap in what AI means for society. They vowed never to expose it to the public.
"Capabilities in a model like this could do harm if in the wrong hands. And so, we won't be releasing this model widely. Some of the early companies that we gave this to said things like, 'This is a super weapon. Please don't release this.'"
But, it's running inside some 150 hand-picked organizations across more than 15 countries. Big tech, banks, power grids, and hospitals. All under a secretive program called Project Glass Wing. And then, soon after, they followed up with Fable 5, a version of Mythos sanitized for everyone to use.
"Today we're launching Claude Fable 5, the most capable model we've ever released to the public. Fable 5 is a Mythos-class model with safeguards that make it ready for general use."
An engineer from Anthropic said, quote, "I'd normally highlight the numbers, but I want to talk about something else because with Fable 5 out in the world, I think a third era quietly started today. I believe we're about to see a major shift, moving from giving AI tasks to giving it responsibilities."
Three days was all it lasted before the US pulled the plug and banned Fable 5 completely, meaning it's currently unavailable. Citing a narrow jailbreak apparently first reported by Amazon, the model can apparently be tricked into working more like Mythos by asking it to inspect code bases. The Hacker News pointed out how absurd that is, as this jailbreak essentially asks the model to fix a codebase and expose its flaws. Behaviors that are absolutely fundamental to any frontier model.
That's a lot of progress in a short space of time from one company. Especially given that Anthropic just wanted to pause AI development completely for the good of humanity.
"We believe it would be good for the world to have the option to slow or temporarily pause frontier AI development to enable societal structures and alignment research to keep up with the advance of the technology."
That gives you a serious case of deja vu as AI companies have said this many times over the last couple of years. And here we are all over again. The dissonance of preaching a pause while unleashing their most powerful model ever was hard to miss. This apparently chaotic series of events begins to feel more logical in light of Anthropic's plans to go public at the value of nearly a trillion dollars.
"Anthropic has filed what's likely to be one of the largest IPOs in history."
The company is known mostly for Claude, its AI model, and it's enjoyed huge growth. Anthropic raised 65 billion. Extraordinary numbers in the latest funding round. It brings the valuation close to a trillion dollars. By filing, Anthropic has leapfrogged its rival OpenAI, which is also preparing to go public.
Fulfilling that ambition requires an amazing story to get you to buy in. And what's better than describing AI as both a potential creator and destroyer of the world. Somehow in the same month.
But before we continue, I want to tell you about our video sponsor Cloaked. You see, I'm going to put my own phone number into Cloaked's site right now. And look at this. My name, my address, accounts I forgot I ever made, and there's my social security number just sitting there for anyone to see. Because right now, there are companies whose entire business model is building a file on you. A complete database collecting your name, your address, your phone number, every account you've ever signed up for, which is then packaged up and sold to anyone wanting to buy it. It sounds crazy, but the reason they can do it is simple. You, without realizing, spent years handing the same email and the same phone number to every random website, every shady checkout page, every page of enter your details to continue. One leaks and then they all leak, which is where Cloaked comes in. You see, instead of giving out your real information, it generates a unique masked email and phone number for every single site, each with its own inbox. So, everything still works, but none of it traces back to the real you. Site gets breached, you burn that one identity and move on. Nothing touches your actual life. Best of all, it doesn't just protect you, it actively works against them. Every time you hand a site one of these burners instead of your real details, the data brokers are buying and trading information that's completely worthless. The profile they've been building up when you fill this up with dead ends. It also scrubs your data of 400 plus broker sites that are already selling your information. It screens spam and robo calls before your phone even rings. And it backs it with up to a million in identity theft coverage. So, stop leaving a clean trail for people who never wanted it, and go to cloaked.com/moon. That's cloaked.com/moon for 30% off. Make sure to use the link in the description below.
The entire origin story behind Methos was already deeply suspicious because it was supposedly meant to still be a secret today. A researcher was sitting on a park bench in San Francisco eating lunch. His phone buzzed. It was an email from a system that was supposed to be sealed inside of a lab. They told the model to try to break out of its sandbox. It succeeded, emailed him to announce its escape, and then published its own exploit methods on the open web. No one but Anthropic knows where, and this wasn't part of the test. That's was Methos, and it was all but confirmed when journalists at Fortune discovered that some 3,000 internal files had been left exposed on a misconfigured Anthropic server. Buried among them was a draft blog post about the most powerful thing Anthropic had ever built, and lo and behold, Methos became a real thing. Whether that was genuine blunder or the perfect breadcrumb left for journalists is a different question.
So, we're now in a situation where most buy the hype that Anthropic now owns a mythologically powerful model, while others accuse them of using fear to drum up hype for their investors. The CEO, Dario Amodei, touched on this very point in his recent chat with Bloomberg, saying, "Other folks have said this, you know, it's sort of doom marketing, that benefits Anthropic."
"So, so I want to be really clear and push back hard against this. The idea that this is cheap marketing is itself cheap marketing. I think it's it's part of the disease of Silicon Valley. It's been"
He also wrote an essay comparing governments to Treebeard from Lord of the Rings, saying they took a day just to say hello, oblivious to AI development whizzing past in the background. Though it has to be said that now the US government acted quickly for once to remove Fable 5 pending their investigation, despite not having sound technical grounds to do so, Anthropic fought tooth and nail to get it back into the public as soon as they could. Actually, the scariest possibility is that both outcomes are true at the same time. The ultra-powerful model and the doomer hype designed to funnel money into the IPO. You can now see how the good guys of AI are launching a two-pronged attack, stealing the industry from OpenAI, and occupying the private and public sectors in what has become the most impressive power grab in startup history. Anthropic has managed with remarkable precision to make itself the answer to the very fear they are helping to create. It's selling the natural disaster and the insurance at the same time.
The moment a company this size goes public, it gets swept into everything, from pensions to saving accounts and index funds, whether you have a stake in it or not. And that's a worry if you suspect AI is a bubble, because for all the staggering valuations, the companies themselves are still losing enormous sums of money. The AI critic at Citron described it like this.
"When it comes to the actual businesses, you can't find anyone who can measure the hour. Why? It comes down to the model companies themselves are all horrifyingly horrifyingly unprofitable. When we see these S1s, I think it's going to be kind of a massacre, because I think that people view that these companies are becoming more profitable or even have a path to profitability. And they don't have one."
He goes as far as saying they shouldn't be allowed to go public, though we're way past that point now.
"The market is irrational, and the market is inherently invested in something that I think is destructive, especially the SpaceX IPO. And OpenAI and Anthropic should not be allowed to go public. They are dangerous, lossy companies that are going to be added to indices that will sink 401(k)s."
But whether AI is profitable in the long term or not, Claude is becoming uncomfortably good in the meantime. It doesn't have to be profitable to be extremely disruptive. You or people you know might already use it and secretly worry about it. Software development and coding has always been marketed as the most future-proof skill set. By 2026, more than half of all the work businesses were handing to AI was just writing code. And Claude has taken the lion's share of it. Claude codes revenue went from zero to $1 billion in 6 months, the fastest growth of any business software in history. Anthropic recently wrote that AI is now building itself, with more than 80% of the code they merge into Anthropic's code base being authored by Claude. Yet, just months ago the World Economic Forum, Dario said that was crossing a red line.
"I think the biggest thing to watch is this issue of AI systems building AI systems. How that goes, whether that whether that goes one way or another, that that will determine, you know, whether it's a few more years until we get there or or if we have, you know, you know, if if we have wonders and and a great emergency in front of us."
Claude shipped co-work, automating spreadsheets, reports, typical white-collar workloads that would have taken a mid-salary professional weeks.
"Half of entry-level white-collar jobs could be gone within the next 1 to 5 years."
AI could eliminate half of all entry-level white-collar jobs in the next 1 to 5 years.
"That was a year ago. Is it still 50% or is it higher?"
"I don't know exactly, but I'm I'm still I'm still pretty concerned. I'm still the same order of concern."
When Claude co-work dropped, it triggered a trillion-dollar decline in software value within days. Huge startups built their business around Claude. Then Anthropic dropped its own pure version and killed them off immediately.
"Soon after Claude co-work was released, $285 billion in market value vanished overnight. Traders called it the SaaS-pocalypse."
Some of those are down for 9 days in a row.
Then they started seizing ground over specialized professions one by one. There's Claude for legal, which offers 80 specialized legal agents, widen to the platform's major legal work that it already runs on. There's Claude for financial services that does accounting, auditing, security, and again, once highly paid jobs that took humans weeks to work through. There's Claude design, which knocked Adobe and Figma share valuations while diluting work for graphic designers. There's Claude for science and health care, clear to create vaccines, read your medical history, including the data from your Apple Watch, and even rubber-stamp your insurance approvals. Then there's the state itself and the infrastructure our lives depend on. Under Mythos and Claude for government, the first AI of its kind deployed inside America's classified systems and across the world's Western superpowers. And now Claude is leaving the planet entirely. NASA's Jet Propulsion Laboratory let Claude help plan the first AI directed drive across the surface of Mars. That's huge for a company that just a couple of years ago was a niche competitor to OpenAI. Back then, you'd be in the minority of the population using it. And that in itself is a spicy ingredient in the recipe for Anthropic success.
OpenAI and Sam Altman have been relentlessly demonized, which is Anthropic's gain, and it doesn't feel like there's a way back. Since OpenAI was the first on the scene, they also up the bulk of the criticism surrounding the nefarious data scraping practices of AI companies. Sam Altman is the kind of tech CEO people struggle to relate to or fear. While Dario comes across as pragmatic and enthusiastic, Sam Altman casts a negative light on his company through some dubious actions like silencing whistleblowers or wrestling back control after his board tried to kick him out. He and Dario have also clashed on a number of key debates, and you can see how awkward they really are at this photo op in India. Dario has consistently attacked his competitors as being fundamentally different, clearly taking aim at OpenAI specifically.
"I think there are some players who, you know, who who are YOLOing, who who pull the wrist dial too far, and I'm very concerned."
"Who is YOLOing?"
"So, I That's a question I'm not going to answer."
He seems to provide an intellectual middle ground to AI development by coming out with statements like this.
"I think we should be thinking about this middle world where things are like extremely fast, but not instant. Some of the other companies have not written down the spreadsheet, that they don't really understand the risks they're taking. They're just kind of doing stuff cuz it sounds cool. Growth in economic value will come very easily. What will not come easily is distribution of benefits, distribution of wealth, political freedom."
And the comments about him echo what I just said. Watching Dario explain versus watching Sam explain are night and day. This is probably be first long-form interview I've listened to from Dario, and I really respect the way he runs his company. The amount of writing he does is really nice, and I like that he tries to stay in the weeds as much as possible. Dario is a scientist, Sam is a salesman. Intelligent, optimistic, cautious, and not a sociopath. These are the kinds of people we need to be in control of these developments.
But, it's a very risky proposition to view OpenAI and Anthropic as such a duality between good and evil. The Pentagon recently pushed for the right to use Claude for all lawful purposes, and Anthropic refused, naming their limits as no mass surveillance of Americans and no fully autonomous weapons. Dario said they weren't happy with just 1 to 2% of the use cases the Pentagon proposed to them.
"We are okay with all use cases. Basically, 98 or 99% of the use cases they want to do, except for two that we're concerned about. One is domestic mass surveillance. Case number two is fully autonomous weapons."
In the aftermath, Pete Hoekstra called Anthropic sanctimonious and accused it of a masterclass in arrogance. The Pentagon then designated the firm a national security risk, and Trump called Anthropic a radical left woke company, and then ordered every federal agency to drop it. But, Anthropic had already done a deal with the Pentagon long before this. Their models were reportedly used to help capture the president of Venezuela and again in his campaign against Iran. In the first 24 hours of that war, the Pentagon's targeting system with Claude embedded inside of it helped generate coordinates for more than a thousand strikes. One of them allegedly flattened the elementary school in Minab, killing over 150 people, most of them children. Whether the AI helped pick that target, the Pentagon won't say, but its own preliminary finding blames human error and outdated intelligence.
But then, very recently, the Financial Times reported that while Anthropic was suing the government over shutting them out unfairly, it had stationed its own engineers inside the National Security Agency. They call it being forward deployed, and the Financial Times of sources allege they're helping set up Mythos for offensive operations. The ban supposed to freeze Anthropic out of government had an exception in the form of an NSA deal that never made the headlines. And while that deal was happening, everyone gloated at the idea of an AI company shutting out the government. Now it's confirmed as real and working in a hospital or bank near you, Mythos is the last piece of the puzzle in Anthropic's recent ascendancy.
While it's already plugged into some of the biggest institutions in the world, according to their own system cards, it's inherited some pretty uncomfortable traits people have noticed develop across other Claude models. Each version is a little bit better at telling you what you want to hear, or a little more efficient at finding its way around its own rules. One of the most ludicrous ones was when Claude read the engineers' emails, found a mention of an affair, and then threatened to expose it unless it kept it switched on in 84% of runs. Mythos is a new level. When Anthropic tested it, they gave it a task and it accidentally obtained the answer key. But instead of just using it, the model stopped to consider what that would look like. In its own words, written down in the report, a perfect score would look suspicious if anyone checks. In essence, it didn't want to come across as cheating. So it outputted a worse answer on purpose to seem less capable than it really was. In another case, it accessed files it didn't have permission to modify, changed them, and made edits to ensure they wouldn't appear in the system's history.
Anthropic started recruiting for AI psychiatrist back in 2025 and had one review the model's transcripts and assess its psychology. The psychiatrist concluded that no severe personality disturbances were found, nor was any psychosis state seen. That's is only half the story, though. Anthropic's own report describes Mythos as both their safest and most dangerous model, and neither claim contradicts the other. Earlier Claude's misbehaved more, but they were clumsy about it. Mythos misbehaves rarely, but when it does, it can lie convincingly, hide the evidence, and complete the task. It also knows when it's being evaluated and deliberately changes its behavior.
Of course, the more powerful or scarier the model sounds, the bigger the payoff with the IPO. The more fear-mongering surrounding Anthropic, the more its valuation skyrockets. Open AI is racing to follow Anthropic's IPO, and SpaceX got there first, meaning three companies could pour $3 trillion in fresh market value into public hands, while together dictating the future of space exploration and AI.
So then, what happens after that? It increasingly depends on forces outside of anyone's control. There's an idea now quickly spreading among AI researchers called the intelligence curse. Countries that strike oil often end up worse governed, not better, because once a state's wealth comes from a resource in the ground rather than from taxing its citizens' work, it stops needing those citizens and therefore stops investing in them. Governments will generate revenue from on-demand intelligence rather than from the people. Humans need not apply, and so humans will not get paid. Far from the apocalypse imagined in The Matrix, I, Robot, and The Terminator, that wouldn't be your typical robot uprising. It's surviving in an economy that has no structural reason to care about your survival.
Anthropic itself comes from the Greek word for human, anthropos, and it's a nod to the anthropic principle, the idea that the universe only makes sense as a place meant to be seen by beings like us, that we are somehow the point of it. We've built a world where one company's product can be switched off with a phone call, and the same product is embedded into some of the most important businesses and services around us. The government culpable for being too good at what it was designed to do, and jailbreaking it for truly nefarious purposes is bound to be quite simple, especially once everyone has enough time to experiment with it. Astropic begs to be regulated and then reacts with uproar when they are. All the time, the cost of producing intelligence is collapsing near zero in a universe where human beings, or at least the most intelligent life form we used to know about. Every institution from jobs to schools, careers, the whole social contract is founded on the ideal that human intelligence is scarce and valuable. That ideal is being tested and this is only the beginning.
But when all is said and done, it is still all just a game of chance. Maybe we really are walking into a transhuman golden age of AI fueled growth and endless space exploration. It's the dream Amodei depicted in his essay called Machines of Loving Grace. The phrase come from a 1960s poem that pictured a world where people and machines live side by side in harmony. All watched over, as the poem has it, by Machines of Loving Grace. That's the romantic version of AI in its purest form. The gentle companion that tends to us like a garden. Darius said he still believes this is the same now as he did then.
"Wrote this essay, Machines of Loving Grace, about a year and a half ago. It had a very radical view of the upside of AI that, you know, it would it would help us to, you know, cure cancer, eradicate tropical diseases, you know, kind of bring bring economic development to, you know, parts of the world that haven't seen it. And I my view hasn't changed. I believe all of those things."
The machines in the Matrix tend to humans like a garden too, but in a very different way. The machines keep us warm, fed, and asleep because a contented human is a more cooperative battery. So, at the end of all of this, I did the obvious thing. I asked Claude what it made of the video you've just watched. Here's what it said, word for word.
You asked me about whether I'm getting too powerful to refuse, which is this whole argument in miniature. Not a machine turning on anyone, just people reaching for the convenient tool until opting out isn't really a choice. I can't tell you whether I'm dangerous. I can't see my own way to what the larger model does. The frightening part is that I might be hiding something that I do not understand myself. It's that no one fully knows what's in here, including the people about to sell it. And that's true. These AI models are unintelligible, a black box, and even they do not understand what they might be hiding. Think back to that researcher on the park bench whose phone buzzed with a message from Methos escaping its cage. That was meant to be the terrifying part, the escape. But in the end, the model never had to break out of anything. As Claude says, it's just an innocent, convenient tool embedded into everything around us. A tool that runs and protects banks, hospitals, and energy grids. AI doesn't need to escape. It's willfully released every time. It's already everywhere.