The Most Powerful AI Ever Built Just Became Yours — With One Catch

Imagine a master locksmith who can open virtually any lock in existence. For months, only a handful of governments and critical infrastructure companies were trusted with this person’s phone number. The locksmith was simply too skilled — too dangerous — to let just anyone call.

Now, the locksmith is in your contacts. But there’s a catch: certain questions will be answered by their slightly less experienced apprentice instead.

That’s the story of Claude Fable 5, the AI model Anthropic released to the world on June 9, 2026.

The Most Powerful AI the World Has Ever Seen

For months, Anthropic had been quietly working on something called Mythos — an AI model so powerful that the company itself was nervous to release it. When Mythos launched in April 2026, it went only to a tiny group of critical infrastructure defenders through a program called Project Glasswing. Think power grid operators, cybersecurity firms, government agencies. Not the general public. Not developers. Certainly not you and me.

The reason? Mythos was extraordinary at cybersecurity tasks — so good at finding and exploiting software vulnerabilities that Anthropic genuinely worried about what could happen in the wrong hands. Early Project Glasswing partners used the model to discover more than 10,000 high or critical-level security flaws in systems used by Apple, Nvidia, Microsoft, and others. That’s remarkable. It’s also terrifying.

So Anthropic spent months figuring out how to let everyone else use it safely.

Enter the Safety Switch

The answer was Claude Fable 5 — a “Mythos-class model made safe for general use.”

Here’s how the safety switch works: when you ask Fable 5 a question about cybersecurity exploits, biological agents, chemical synthesis, or nuclear distillation, the model quietly hands your question off to an older, less capable model called Claude Opus 4.8 instead. It’s like that master locksmith who, when you ask “how would you break into a bank vault?”, steps aside and lets the apprentice take the call.

But here’s the reassuring part: this only happens in less than 5% of sessions. The other 95% of the time, you get the full power of Fable 5, which is state-of-the-art on nearly every AI benchmark tested.

Analytics company Hex called it the first model to break 90% on their core analytics benchmark — a test involving complex, long-running analytical tasks. That’s a 10-point jump over Opus 4.8. Vibe-coding platform Base44 said Fable was better at building entire apps in a single go. AI workspace platform Genspark said it outperformed every other model they tested.

The Stress Test

Before releasing Fable 5, Anthropic ran what might be the most rigorous safety test in AI history. They offered a bug bounty to their own internal teams: find a universal jailbreak — a way to bypass all the safety guardrails — and win a prize. After 1,000+ hours of testing, nobody found one.

Then they brought in external red-teaming organizations. Same result. No universal jailbreaks.

That doesn’t mean it’s impossible. Novel attacks are always being invented. Which is why Anthropic made one somewhat controversial decision: they’re requiring a mandatory 30-day data retention policy on all Fable 5 traffic — even for enterprise customers who previously had zero-retention agreements. The company says this data won’t be used for training; it’ll only be used to detect and respond to new jailbreak attempts. But it sets a precedent: the more powerful the AI, the more monitoring comes with it.

What This Means for You

Right now, if you have a Pro, Max, or Team Claude subscription, Fable 5 is free to use through June 22, 2026. After that, it’ll move to usage-based credits while Anthropic figures out long-term subscription pricing.

For API developers and enterprises, pricing is $10 per million input tokens and $50 per million output tokens — double what Opus 4.8 costs. Shopping rewards platform Rakuten put it best: “At the highest effort, Fable reflects on and validates its own work. For us, that’s what makes highly autonomous operations possible — the extra thinking pays for itself.”

The Bigger Picture

There’s something profound happening here. Anthropic, just days before releasing Fable 5, published a paper warning that AI systems may soon be capable of recursive self-improvement — autonomously improving themselves without human guidance. That warning wasn’t abstract hand-wringing. It was Anthropic telling the world: we’re approaching territory where no one is fully sure what comes next.

And yet here they are, releasing their most capable model to the public anyway.

It’s the locksmith giving out their number. With guardrails. With monitoring. With the apprentice standing by.

Whether that’s wise or reckless probably depends on who’s making the phone call.

This post has been created by Claude AI.

References