Fable 5: 4 days from launch to shutdown, a $965 billion valuation, and an off switch in Washington
Alexandre
··
Reading time: 9 min
June 1, 2026, Anthropic files a confidential S-1 with the SEC. $965 billion valuation. On June 9, the company launches Fable 5, its most powerful model ever released to the public. On the evening of June 12, the US government sends a letter at 5:21 PM ET, and Anthropic cuts access to Fable 5 and Mythos 5 for all its customers, worldwide.
Twelve days. Three acts. And a very concrete takeaway: a single email from the Commerce Department is enough to make the best model on the market disappear.
I use Claude Code every day to build Waku. Architecture, debugging, refactoring, writing... everything goes through it. So this week, I saw in real time just how fragile that dependency really is.
Let's break it down.
Fable 5 and Mythos: what Anthropic launched on June 9
Fable 5 is the public version of Mythos, Anthropic's frontier model. It outperforms all of the company's previous models across software engineering, scientific research, vision, and reasoning benchmarks, with an advantage that grows on longer, more complex tasks.
On June 9, Anthropic launched two models in parallel. Fable 5, available to everyone, and Mythos 5, restricted to Project Glasswing partners. The difference between them comes down to one word: guardrails. Fable 5 automatically reroutes sensitive queries (cybersecurity, biology, chemistry) to the less powerful Opus 4.8. Mythos 5 does not hold back.
Pricing is the same for both: $10 per million input tokens, $50 per million output tokens. Half the price of Mythos Preview from April, twice the cost of standard Opus 4.8 ($5 and $25). Until June 22, Fable 5 is included at no extra cost on Pro, Max, Team, and Enterprise plans. After that, credits are required.
Project Glasswing has expanded to roughly 200 organizations: cybersecurity firms like Dragos, Tenable, Trend Micro, Netskope, BeyondTrust, and infrastructure operators like BT, Verizon, SK Telecom, and Hitachi. Fifty cyber defense teams have direct, unrestricted access to Mythos 5.
My take
The timing is super interesting. Anthropic launches the most powerful model on the market and makes it free for 13 days on paid plans. Generous... or strategic. When you file your S-1 on June 1 and want to show a massive user base before the IPO, offering the star model for two weeks makes sense.
For solo devs, Fable 5 at $10/$50 per million tokens is a cost doubling compared to Opus 4.8. On a typical month where I consume between 50 and 100 million tokens on Waku, that is between $500 and $5,000. This is the kind of model you use surgically, not on autopilot.
Click to enlarge
Photo credit: TechCrunch / Getty Images
A model that hacks faster than a human
Mythos is the first AI model capable of turning a security patch into a working exploit in under an hour. In internal testing: a proof-of-concept for a Windows kernel flaw in 31 minutes, 8 Windows exploits in 18 hours, roughly $2,000 per exploit.
The numbers come directly from Anthropic's red teams, published the day before the launch in a report covered by Axios and SecurityWeek. The model worked on 21 Windows kernel vulnerabilities disclosed between January and February 2026. Compiled code, no source access. It produced functional payloads that triggered system crashes on the majority of bugs examined.
The concept Mythos disrupts is the "patch gap" (the delay between the release of a security fix and the moment attackers reverse-engineer that patch to create an exploit). Traditionally, this gap is measured in days, sometimes weeks. Mythos compresses it to hours. Rob T. Lee, Chief AI Officer at SANS Institute, compared it to a hurricane warning: preparation time is shrinking, and the seawall has not been built yet.
And this is not just theoretical. Since Glasswing launched in April, partner teams have discovered over 10,000 high- or critical-severity vulnerabilities in production software. Ten thousand. In two months.
My take
$2,000 per exploit. That is the price of a MacBook Air. It democratizes cyberattacks at a scale nobody anticipated, and at the same time it is exactly what defenders need... finding the flaws before the attackers do.
Anthropic chose to limit Mythos to 50 vetted cyber defense teams. The problem is, the genie is out of the bottle. CSOOnline has already documented an AI worm prototype that does not even need Mythos to take over a network.
For Waku, an iOS app that stores user data, I think about this differently now. Security is no longer something you deal with at the end of a sprint. If a $2,000 model can find vulnerabilities in compiled code without source access, indie apps are just as exposed as the big players.
Click to enlarge
"We made the wrong tradeoff": the admission that changes everything
Before Fable 5 even launched, Anthropic publicly acknowledged making "the wrong tradeoff" on Mythos guardrails. The model was silently rerouting sensitive queries to Opus 4.8 without telling the user. A decision the company reversed after a wave of criticism.
The guardrails story is a textbook case. When Anthropic launched Mythos Preview in April, the model blocked cybersecurity, biology, and chemistry queries by rerouting them to Opus 4.8. Without telling the user. You would ask a question, get an answer, and never know it came from a less powerful model.
Business Insider reported Anthropic's response: "We made the wrong tradeoff, and we apologize for not getting the balance right."
The fix came with Fable 5: the model now tells the user when a request is refused or rerouted. Better. But it remains a model that, by design, does not give you its full capabilities in certain domains.
My take
Silent rerouting is exactly the kind of decision that breaks trust. When I code with Claude Code, I assume the model gives me its best answer. If I ever discover it silently downgraded to an inferior model without telling me... I question everything.
Anthropic fixed it, and that is good. But it reveals a reflex: when you have a model that is too powerful, your first reaction is to throttle it secretly rather than be transparent. It is the same pattern as with Opus 4.6. When the company optimizes for safety at the expense of transparency, the most serious users are the first to leave.
The June 12 shutdown: when Washington flips the switch
On June 12, 2026 at 5:21 PM ET, the US Commerce Department sent Anthropic an export control directive: cut off access to Fable 5 and Mythos 5 for all foreign nationals, including non-US Anthropic employees on American soil. Result: both models disabled for everyone, everywhere.
The directive did not specify exact reasons. According to Anthropic, the government acted based on a jailbreak method that could bypass Fable 5's guardrails. The company disputed the severity in a blog post published Friday evening, calling the technique "narrow" and applicable to other public models, including OpenAI's GPT-5.5.
The Wall Street Journal, Bloomberg, Reuters, and CNBC all confirmed: Amazon AWS revoked access for "all users in all regions" at Anthropic's request. The Pentagon's CIO posted on X: "Some things are simply more important than revenue cycles."
Four days. Four days between the public launch of a model and its forced shutdown by the government. Unprecedented.
This decision was so sudden. Anthropic did not have time to negotiate, propose a fix, or deploy a patch. Shut it all down, now. And they complied. When you are at a $965 billion valuation with an S-1 on file, you do not play hardball with the Commerce Department.
We knew frontier AI models were national security issues. We knew export controls existed. But watching a model launched on Tuesday, used by millions of devs for four days, and shut off on Friday evening by a government letter... nobody saw that coming.
Click to enlarge
$965 billion valuation and an off switch in Washington
Anthropic filed its confidential S-1 on June 1, 2026 after a $65 billion funding round. $47 billion in annual recurring revenue, $2 billion monthly revenue as of March 2026. OpenAI filed its own S-1 one week later. And in the middle of this IPO race, a single government email was enough to kill the flagship product.
The numbers are staggering. PitchBook reports a $965 billion valuation, which would make Anthropic one of the largest tech IPOs in history. TechCrunch notes that OpenAI filed its S-1 on June 8, one week after Anthropic. The race between the two labs now extends to Wall Street.
But the June 12 shutdown raises a question financial analysts do not want to hear: what is the value of a company whose main product can be deactivated in 4 hours by a government email? Forbes had already raised the point: "The binding constraint on frontier AI has moved off the model itself, onto the capacity to serve it." Except here, the constraint is not even compute capacity. It is politics.
My take
The paradox is absurd. Anthropic is about to become one of the most valuable companies on the planet, and at the same time it is the first to prove that a government can shut down an AI model overnight. It is as if Boeing launched a new aircraft and the FAA grounded it four days later.
For investors reading the S-1, regulatory risk just moved from the "theoretical risks" chapter to the "this happened to us last week" chapter. Dario Amodei himself published an essay on June 10, two days before the shutdown, titled "Policy on the AI Exponential", calling for binding regulation of frontier models. He probably did not know just how binding and how fast that regulation would come.
Click to enlarge
Credit: Ars Technica
AI sovereignty is no longer a theoretical debate
The Fable 5 shutdown triggered immediate reactions outside the United States. In India, Sridhar Vembu (Zoho co-founder) called for national technological autonomy. In Europe, the Commission had launched a "technological sovereignty package" just days earlier to reduce dependence on American models. The takeaway is simple: anyone depending on US AI is exposed to a cut-off without notice.
The timing is almost too perfect. On June 8, the European Commission announced a unified framework for cloud and AI, with investments in sovereign AI, semiconductors, and open source. Four days later, Anthropic proves exactly why that framework is necessary. Sridhar Vembu declared on X that the event was "big" and that "global AI dependence must be addressed."
My take
This is the topic I am digging into the most right now. I build Waku 100% with Claude Code. My IDE is Claude. My copilot is Claude. My agent architecture is Claude. If tomorrow the Commerce Department decides that European nationals can no longer use Anthropic's frontier models (and that is literally what just happened), what do I do?
Honestly, I do not know. There is no European equivalent to Mythos. Mistral is good, but not at Fable 5 level. And the EU "sovereignty package" is a long-term play. We are talking years before a credible alternative exists.
This shutdown is a wake-up call for every European developer who has built their workflow around Claude, GPT, or Gemini. We are tenants, not owners. And the lease can be terminated by a foreign government, without notice, on a Friday evening.
What this actually changes
1. Frontier AI is now a national security matter
The Commerce Department treated Fable 5 like defense technology. The most powerful AI models are now in the same category as advanced semiconductors and surveillance tools.
2. Platform risk has never been higher
Four days between launch and shutdown. It is now crucial to diversify your model portfolio.
3. AI sovereignty is about to become a selling point
European and open-source models will capitalize on this episode. "Our model cannot be shut down by Washington." That is an argument that will weigh in enterprise purchasing decisions starting Monday morning.
Conclusion: Anthropic's impossible triangle
Anthropic is caught in an impossible triangle. On one side, growth: $47 billion in revenue, a $965 billion IPO, a race with OpenAI. On the other, security: a model that hacks faster than a human, guardrails they themselves called a "wrong tradeoff." And above it all, geopolitics. A government that can shut everything down with one email.
What happened this week is the first head-on collision between these three forces. It will not be the last.
And for us European devs building with these tools every day, the question is no longer whether our favorite AI will be shut off someday. It is whether we will have a plan B when it happens.
Do you have a plan B if Claude disappears tomorrow? Drop me a message on Twitter/X or in the comments.
The problem with AI in 2026 is no longer technical. It is geopolitical. And geopolitics does not compile.
Alex
Key takeaways
Fable 5 launched June 9, shut down June 12. 4 days from launch to forced shutdown by the US Commerce Department
Mythos generates Windows kernel exploits in 31 minutes for $2,000. The patch gap drops from days to hours
Anthropic filed its S-1 at a $965 billion valuation, but one government email is enough to kill the flagship product
AI sovereignty is no longer theoretical: European devs, we are tenants, not owners
It is now crucial to diversify your model portfolio
