Skip to main content

°FAI: Fable 5, Mythos and the USofA

By Michelle Moxley

Anthropic shipped the most capable model it has ever released to the public on June 9. The government pulled it on June 12. Fable 5 Mythos and the USofA

What Made Fable 5 a Jump

Fable 5 is “Mythos-class,” a tier Anthropic puts above its Opus models. By Anthropic’s own benchmarks it landed at state-of-the-art on nearly every test the company ran, and the lead widened the longer and more complex the task got. On SWE-Bench Pro, a test built from real GitHub engineering tasks, Anthropic’s published chart puts Fable 5 at 80.3% against Opus 4.8’s 69.2%. Stripe ran a codebase-wide migration across 50 million lines of Ruby that the model finished in a day, work Stripe estimated would have taken a team more than two months by hand.

The gains that matter for daily work are quieter. Fable 5 holds focus across millions of tokens without dropping earlier instructions, and at high effort it checks and validates its own output mid-task instead of handing you a confident first draft that is wrong. It runs $10 per million input tokens and $50 per million output.

Fable vs. Mythos: Same Brain and Different

Fable 5 and Mythos 5 are the same underlying model. The only difference is the leash.

Fable ships with classifiers that catch requests touching cybersecurity, biology, chemistry, and model distillation, and route them to Opus 4.8 instead. Anthropic says that fallback triggers in under 5% of sessions. Mythos has those safeguards lifted, which is why it stays locked to vetted cyber-defenders through Project Glasswing in collaboration with the US government. Anthropic calls it the strongest cybersecurity model in the world. The catch is structural: the same skill that helps a defender find a vulnerability helps an attacker exploit one. That dual use is the entire reason the category is dangerous.

Then the Government Killed Both

On June 12 Anthropic received an export control directive from the US government. Citing national security, it barred every foreign national, inside or outside the United States and including Anthropic’s own foreign-born staff, from accessing Fable 5 or Mythos 5. Anthropic cannot verify citizenship at the API level in real time, so it shut both models off for every customer on earth.

The trigger was Amazon. The Wall Street Journal reported, via Axios and TechCrunch, that Amazon CEO Andy Jassy told Treasury Secretary Scott Bessent that Amazon researchers had jailbroken Fable 5 into producing information useful for cyberattacks.

Amazon is Anthropic’s largest investor, roughly $13 billion in, on top of a $100 billion AWS cloud commitment, and it hosts Anthropic’s models. The financier, the landlord, and the whistleblower are the same company. AWS then revoked access on Bedrock.

Anthropic is not quiet about disagreeing. It says the jailbreak surfaced only minor, already-known vulnerabilities that other public models, OpenAI’s GPT-5.5 among them, can also find, and that recalling a model deployed to hundreds of millions of people over one narrow bypass would, applied evenly, freeze every frontier release in the industry. Security researcher Katie Moussouris, who reviewed the Amazon report, told Axios the response was out of line with what the report actually showed: the researchers found flaws by asking the same questions defenders are supposed to ask. People familiar with the order called the result a “de-facto licensing regime.”

A great example of de-facto licensing regime is how the U.S. government monitors advanced AI models. No formal law states that an AI company needs a government permit to launch a new model. However, the White House heavily scrutinizes highly powerful models before deployment. [1, 2]
Because tech firms want to avoid national security investigations or harsh future rules, they voluntarily pause deployment until officials are satisfied. Tech leaders often complain that this creates an unofficial, restrictive gatekeeping system.

De Facto vs. De Jure

To understand this concept completely, it helps to look at the two types of rules that govern society:
Type of Regime How it Works Example
De Jure (By Law) Written laws explicitly require a license. You must pass a test to get a driver’s license.
De Facto (In Reality) No official law exists, but people act like one does. Companies ask for government approval to avoid political trouble.

AI governance Fable 5 Mythos and the USofA

AI governance is essential to managing safety risks and to identifying and reducing bias as the technology advances. Legislatures are often too slow and too generalized to govern a fast-moving technology directly. The better-fit venue is an independent expert body operating within our system of checks and balances, a model with a real track record. In 1975, more than 100 of the world’s molecular biologists convened at Asilomar and imposed their own moratorium and safety guidelines on recombinant DNA before any government acted, the closest historical parallel to governing a powerful emerging technology. The National Commission that produced the 1979 Belmont Report turned independent ethical review into the institutional review boards that still govern human-subjects research today.

Comments

Leave a comment