Anthropic just gave everyone its most powerful AI, with the dangerous bits locked away

For months, Anthropic's Mythos model was the AI you weren't allowed to touch. It launched in April as a preview limited to a handful of partners, mostly because of cybersecurity worries. This week, a version of it landed in everyone's hands.

It's called Claude Fable 5, and Anthropic says it beats every model the company has ever released publicly, with state-of-the-art results in software engineering, knowledge work and vision. Hacker News spent the day arguing about it, and it topped the tech headlines within hours.

The same brain, with a lock on certain doors

Here's the unusual part. Fable 5 and Mythos 5 are the same underlying model. The difference is the guardrails. Ask Fable 5 about high-risk topics in cybersecurity, biology or chemistry and it won't answer directly. Instead it quietly hands the question to Claude Opus 4.8, the older model, and you get that response instead.

Anthropic says this happens in under 5 per cent of sessions. The unrestricted Mythos 5 goes only to approved organisations, mainly cyberdefenders and critical infrastructure providers across 15 countries.

Before launch, Anthropic ran a bug bounty that produced no universal jailbreaks in over 1,000 hours of testing, then brought in external red teams who also came up empty.

What it costs

Pricing is $10 per million input tokens and $50 per million output, double the price of Opus 4.8. Subscribers on Pro, Max, Team and Enterprise plans get it included until 22 June, after which it moves to usage credits while Anthropic builds out capacity.

There's a sting in the tail for big customers. Anthropic now requires 30 days of data retention on all Fable 5 and Mythos 5 traffic, even for enterprises that previously had zero-retention agreements. The company says the data won't be used for training, only to spot novel jailbreaks and attacks.

Why it matters

This launch is a preview of how the most capable AI may be distributed from now on: one model, multiple tiers of access, with the most dangerous abilities reserved for vetted organisations. Anthropic itself recently urged rival labs to agree a coordinated brake pedal on frontier AI development, warning that systems may soon be able to improve themselves without human help.

Releasing your scariest model days after saying AI is getting too dangerous is quite the move. The mandatory data retention could also set a precedent across the industry, where access to top-end models comes bundled with surveillance of how you use them, framed as a safety measure.

And with Anthropic heading for the public markets alongside OpenAI and SpaceX, the timing is no accident. The most powerful AI you can buy is now also a sales pitch to investors.