Anthropic Just Split Its Smartest Model in Two: Meet Mythos 5 and Fable 5

Most model launches are a number going up. This one was a number going up and then splitting in half.

On June 9, 2026, Anthropic released Claude Mythos 5 and Claude Fable 5 - and the interesting thing isn't that they're the most capable models the company has shipped. It's that they're the same model, released under two very different sets of rules.

One brain, two leashes

Mythos 5 and Fable 5 share the same underlying intelligence. The difference is what happens when you ask them something dangerous. Fable 5 ships with a layer of safety classifiers that watch for prompts touching biology, chemistry and cybersecurity. When one trips, Fable 5 doesn't refuse outright - it quietly steps down to the less capable Opus 4.8 to answer. Mythos 5 has no such fallback, which is exactly why it's distributed under strict, limited access.

That's a notable design choice. Instead of a single model that sometimes says "I can't help with that," Anthropic built a public model that, on sensitive ground, simply gets dumber on purpose. It's a graceful-degradation strategy borrowed from systems engineering and pointed at safety.

Why the caution?

Because Mythos-class models are good enough at the wrong things to make people nervous. The capability that keeps coming up is cybersecurity: Mythos 5 is genuinely strong at reasoning about software vulnerabilities - strong enough to find weaknesses in code that was assumed to be secure. That's a gift to defenders running audits and a obvious hazard in the wrong hands, which is the entire reason the unrestricted version isn't sitting behind a public signup form.

Anthropic paired the launch with real-time cybersecurity safeguards: certain high-risk requests now lead to refusals at the model level, not just in the docs. It's the clearest sign yet that the frontier labs have stopped treating "what could this be misused for" as a footnote.

What it means if you're just trying to ship

For the vast majority of developers, Fable 5 is the headline. It's the first time Mythos-class intelligence has been available to the public at all, and for ordinary work - coding, analysis, writing, agentic tasks - you'll likely never see the safety fallback kick in. You just get a noticeably sharper model.

• Writing is the surprise. Testers keep describing Fable 5's prose as clearer and warmer, with fewer of the tics that make AI text feel like AI text.
• Agentic work is the headline. Long, autonomous runs that used to need babysitting now finish more of the time without a human nudging them back on course.
• The leash is mostly invisible - unless your work legitimately lives near security or bio/chem, in which case you'll feel Fable 5 hand off to Opus 4.8 and lose a step.

The bigger pattern

The Mythos/Fable split is a preview of how frontier models probably get released from here on out: not one model for everyone, but a powerful core wrapped in different amounts of restraint depending on who's asking. The benchmark chart will tell you Mythos 5 is the best model Anthropic has made. The release strategy tells you something more interesting - that the labs now consider "who gets the unrestricted version" to be part of the product.

Want to see how Fable 5 stacks up on context window and price against everything else shipping right now? It's in our model comparison, sitting at the top of the Anthropic stack.