The company drew a line the government wanted erased.
In a moment that captures the central paradox of our technological age, Anthropic has released its most capable artificial intelligence to the general public while deliberately withholding its full potential — a company choosing to limit its own creation out of fear of what it might enable. Fable 5, the first public model from Anthropic's advanced Mythos class, arrives not as an open door but as a carefully guarded threshold, where the same system that answers questions also decides which questions deserve answering. The launch unfolds against a backdrop of government entanglement, financial pressure, and a principled refusal to arm power without condition — a rare instance of a technology company drawing lines it is willing to lose business over.
- Fable 5 is genuinely more capable than anything Anthropic has released before — excelling at code, research, and image analysis — but its most powerful features are deliberately locked behind a routing system that redirects sensitive queries to a weaker model.
- The restricted version, available only to roughly 200 vetted organizations across 15 countries including the Pentagon and the White House, can reportedly identify vulnerabilities in critical infrastructure like power grids and banking systems — capabilities serious enough that outside experts spent over 1,000 hours trying to break through the safeguards and failed.
- Pricing signals where Anthropic sees the stakes: at double the cost of its previous model, Fable 5 is an expensive bet that power justifies the premium, even as the company burns through $1.25 billion per month leasing computing capacity from Elon Musk's xAI.
- Anthropic's refusal to lift restrictions on mass surveillance and autonomous lethal weapons has already cost it Pentagon contracts — and the company is now navigating an IPO while locked in a standoff with the Trump administration over where the boundaries of AI power should lie.
Anthropic released Fable 5 to the public this week — its most advanced model yet, and the first from the Mythos class to leave the company's controlled environment. When Mythos was first announced in April, Anthropic kept it entirely restricted, citing cybersecurity risks serious enough to make executives reluctant to let anyone near it. The public release now comes with a significant caveat: the model's most sensitive capabilities remain off-limits.
The mechanism is quietly elegant. When a user asks Fable 5 about cybersecurity vulnerabilities, biological or chemical weapons, or attempts to extract its underlying architecture for use in authoritarian contexts, the system silently reroutes the query to Opus 4.8, a less powerful model released in late May. Anthropic is, in effect, using its own AI as a gatekeeper for itself.
A fully unrestricted version exists, but only for a curated group of around 200 organizations — governments and companies that have already demonstrated they can be trusted with it. The Pentagon and the White House are among them. The government's interest is not incidental: Anthropic says Mythos can identify weaknesses in critical infrastructure, the kind of systems that define national security. To stress-test the restrictions, the company brought in outside experts for more than 1,000 hours of adversarial testing and ran a public bounty program for anyone who could unlock the model entirely. No one did.
The cost of access reflects the weight of the technology. Fable 5 runs at double the price of its predecessor — $10 per million input tokens, $50 per million output — a premium Anthropic is betting the market will absorb. The company itself remains unprofitable despite rapid revenue growth, and is currently leasing computing infrastructure from Elon Musk's xAI at $1.25 billion per month.
The launch lands in complicated territory. Both Anthropic and OpenAI announced IPO intentions within the same week, signaling a new phase of financial ambition across the industry. But Anthropic's relationship with the U.S. government has frayed. The company refused to remove restrictions on mass surveillance and autonomous lethal weapons — even when doing so would have preserved Pentagon contracts that once made Anthropic's tools the only AI systems with defense security clearance. That relationship is now severed. Fable 5 enters the world as a powerful, carefully bounded tool, released by a company that has decided some lines matter more than the contracts that cross them.
Anthropic opened its most powerful artificial intelligence model to the general public this week, but only after building a fence around its most dangerous capabilities. The model, called Fable 5, represents the first public release from the company's Mythos class—a tier of technology so advanced that Anthropic kept it locked away when it first unveiled the lineup in April, citing cybersecurity risks that made executives nervous enough to restrict who could touch it.
Fable 5 is genuinely capable. The company says it excels at writing and debugging software code, wrestling with complex research questions, and analyzing images. For programmers and researchers, it represents a meaningful leap forward. But Anthropic is not handing over the keys entirely. When someone asks the model about cybersecurity vulnerabilities, biological weapons, chemical synthesis, or attempts to extract its underlying technology for use in authoritarian countries, the system routes the query to a weaker model instead—Opus 4.8, which became public in late May. The company is essentially using its own technology against itself, creating a kind of traffic cop that decides which questions get the powerful version and which get the safe version.
Meanwhile, an unrestricted version of Claude Mythos 5 remains available only to a curated list of organizations: companies and government agencies that have already proven they can be trusted with it. That group expanded in early June to around 200 organizations spread across more than 15 countries, and Anthropic expects it to grow. The Pentagon is among the testers, as is the White House, which has established a formal arrangement to evaluate the most advanced models from leading AI companies before they reach the public. The government's involvement reflects a genuine concern: Anthropic says Mythos can identify vulnerabilities in critical infrastructure—banking systems, power grids, the kind of targets that matter to national security.
When Anthropic first announced Project Glasswing, the program that governs access to these restricted models, some observers dismissed the security warnings as hype designed to generate attention. But companies that have actually tested Mythos have since validated the concern. The capabilities are real. Still, the company hired outside security experts to spend more than 1,000 hours trying to break through the restrictions—a process called red-teaming—and ran a bug bounty program offering money to anyone who could find a way to unlock the model completely. No one succeeded.
The pricing tells its own story about where Anthropic sees the value. Fable 5 costs $10 per million input tokens and $50 per million output tokens—double what Opus 4.8 costs. For context, a programmer working intensively on code can burn through a million tokens in hours or less. The company is betting that the power justifies the expense, even as it remains unprofitable despite exponential revenue growth. Anthropic is now leasing computing power from Elon Musk's xAI at a cost of $1.25 billion per month, a figure that underscores how expensive it has become to build and run these systems.
The launch arrives amid a broader financial frenzy around artificial intelligence. Both Anthropic and its rival OpenAI announced IPO plans within the past week. But Anthropic's path forward is complicated by an ongoing standoff with the Trump administration. The company has refused to remove restrictions on mass surveillance and autonomous lethal weapons—lines it will not cross even for government contracts. That principled stance came at a cost: the Pentagon severed its contracts with Anthropic, ending a relationship in which Anthropic's tools were the only AI systems to hold defense security clearance. Fable 5 launches into this tension, a powerful tool released to the world while its maker remains locked in a dispute with the government over what power should be allowed to do.
Notable Quotes
Anthropic says Fable 5 is exceptionally good at writing and debugging software code, answering complex research questions and analysing images.— Anthropic
The company has restricted access to Mythos on cybersecurity grounds, given what it calls the model's ability to quickly identify vulnerabilities in critical infrastructure, including banking platforms and power grids.— Anthropic
The Hearth Conversation Another angle on the story
Why restrict the most powerful version at all? If it's safe enough to release, why not just release it?
Because Anthropic believes some questions shouldn't be answered by the most capable version. A cybersecurity researcher asking about banking vulnerabilities gets routed to a weaker model. It's not about whether the model is safe—it's about controlling what it's used for.
But doesn't that just slow down the person asking the question? They could use the restricted version elsewhere.
True. That's the real tension. The restrictions work only if you trust the people using it, or if you're willing to accept that some legitimate researchers will be inconvenienced. Anthropic is betting on both.
The Pentagon tested it and then cut ties. What happened there?
Anthropic refused to remove restrictions on lethal autonomous weapons. The company drew a line the government wanted erased. So the Pentagon walked away, even though Anthropic's tools were the only ones with security clearance.
Is that a business mistake?
Depends on what you value. Anthropic lost a major contract. But it kept its restrictions. The company is betting that being principled about what its technology can do is worth more than Pentagon money.
What about the price? Double the cost of the previous model—who can afford that?
Companies with serious use cases. Programmers, researchers, enterprises. The price signals that this is specialized, powerful, and expensive to run. It's not for casual use.
And the red-teaming—1,000 hours trying to break it. Does that prove it's secure?
It proves no one found a complete bypass. But security is never finished. It just means the restrictions held under the testing they ran.