SpaceRock

Just a couple of days passed when Uncle Sam pulled the plug on Anthropic’s top models. Gone by then, they’d barely had time to settle in.

Fable 5, along with Mythos 5, went live June 9. Come Friday night, just three days later, they’d vanished everywhere – shut down without warning. Then Anthropic started drafting something sharp, likely the snappiest public jab one top AI group’s made at its own government. Orders arrived from D.C., tucked under export rules, flagged for national security. But according to Anthropic, that reasoning feels shaky. Officials in Washington aren’t saying more.

Let’s start with when things happened, since that part stands out. In March, I published an article at SpaceRock titled Silicon Valley’s New Battlefield: AI, Ethics and the Pentagon, covering how Anthropic declined a Department of Defence agreement valued at nearly $200 million – they refused to work on weapons or monitoring tools. That very week, OpenAI agreed to something like. Saying no drew criticism from federal officials toward Anthropic. By June, though, that same government pulled the plug on Anthropic’s main offerings. It doesn’t mean one caused the other. Still, you’d have to be looking away not to see them together.

Still. The models really did stand out; that’s why seeing it all unfold feels so aggravating.

Fable 5? That’s just what Anthropic names its Mythos-class version, packed with safety filters – ready for anyone to try. The one without quite so many checks goes by Mythos 5, handed only to cyber defence teams through a quiet setup known as Project Glasswing. Same engine underneath, different access rules. I wrote an article on that too. Pricing lands at ten bucks a million inputs, fifty for outputs – less than half the old Mythos Preview fee.

It jumped ahead on tests. Top spot in Cognition’s FrontierCode ranking, hitting high marks even without a full push. Stripe tested it on a massive Ruby project – fifty million lines – and watched it complete a whole-system update in one day, something people would need weeks to do. The toughest logic puzzles from Hebbia’s finance challenge? It scored higher than any other system ever has. First to hit above 90%, Hex claimed, marking a solid leap past Opus 4.8 by ten points on its main test. Over at Cursor, boss Michael Truell noted how it now leads in CursorBench performance, tackling extended challenges that left earlier systems behind.

Right now, chatter on AI Twitter circles mostly around Pokémon. Previous versions of Claude failed to play FireRed, even with extra tools provided to them. But Fable 5 cracked it – just screenshots, minimal code, nothing fancy. When tested on Slay the Spire, memory made all the difference; gains tripled compared to what Opus 4.8 managed under identical conditions. Reaching the last act? That happened far more regularly, too. From a scientific standpoint, Mythos 5 seems to have accelerated certain stages in Anthropic’s internal drug development process roughly tenfold while delivering promising leads for nine out of fourteen protein targets currently under study. When tested without knowing which model generated what output, Anthropic researchers favored the biological insights proposed by Mythos 5 over those from Opus-level systems nearly four times out of five.

That marked the beginning. By the third day, everything had already wound down.

That order landed at 5:21 p.m. ET, June 12, per Anthropic. Around then, they were told to block every foreign national – wherever located – from using Fable 5 and Mythos 5, even their own team members outside the U.S. Shutting down global access became the only real option, since slicing it neatly wasn’t possible.

Not much detail made it into the letter. According to Anthropic, the government thinks someone cracked Fable 5 open. The company looked over what they believe is the real proof behind that idea – turns out it involved having the model scan code and highlight bugs.

Something models do all the time these days for people studying system safety. Those weak spots the model flagged? Already listed online, nothing serious. GPT-5.5 and similar tools floating around can spot them too, no special access needed.

What stands out most is how clear Anthropic’s message feels. Not even the UK AI Safety Institute managed to crack Fable 5 – nor did outside testers after more than a thousand hours hunting flaws. A third-party red team tried thirty known methods meant to force bad outputs, yet saw refusal each time when asking for dangerous hacking help. But here lies the real weight. Pulling a live product just because one limited exploit appears – an exploit that brings nothing new – means no advanced system would survive long online. For any firm building such tech, stability becomes nearly impossible. Failure shows up in small ways with each system. What Anthropic seems to suggest, give or take, isn’t truly about Fable 5 when people demand it pass tests no other tool could clear.

What really lies behind this decision remains unspoken in official circles. On first glance, the action seems sharp yet reasonable enough. When seen alongside the earlier report in March, the picture shifts entirely. A firm that declined Pentagon work – then faced public backlash – suddenly finds its main product blocked over contested reasons. Meanwhile, OpenAI, which agreed to collaborate and placed staff within secure government systems, moves without drawing attention.

For now, Anthropic follows the rules even as it tries to regain entry. The speed of progress – alongside the shape of the outcome – might quietly reveal how much real effort goes into ethical AI by 2026.