“Mythos-Class” AI, Explained Without the Hype

On June 9, Anthropic shipped Claude Fable 5. Buried in the launch was something bigger than a new model. A whole new capability tier they call "Mythos-class," sitting above Opus, which until that morning was the top shelf. (Anthropic's announcement.)

Two names, one brain. That part tripped everyone up at first, so let me lay it out plainly.

• Fable 5 is the model anyone with API access can use.
• Mythos 5 is the same underlying model with the extra safety measures stripped off. Only approved organisations get it, think cyberdefence teams and infrastructure providers.

Same brain, different guardrails. Nothing more mysterious than that.

The reason for two versions

Anthropic says Mythos 5 has the strongest cybersecurity capabilities of any model on earth. You don't hand that to everyone with a credit card. So the public version, Fable 5, ships with safeguards that block responses in high-risk domains like offensive cyber and biology. The clever bit is that those safeguards are the whole reason a broad release happened at all. Strip the guardrails and this model never gets a public API. Expect every frontier lab to copy that playbook.

What it's actually good at

Anthropic claims state-of-the-art on nearly everything they tested. Software engineering, knowledge work, vision, scientific research. The line that matters for our clients is duller and more useful. Both models can work autonomously for longer than any Claude before them. Longer autonomous runs mean agents that finish the job instead of stalling halfway and waiting for a human to notice.

The price tag

$10 per million input tokens, $50 per million output. Double GPT-5.6 Sol ($5/$30) and more than triple Kimi K3 ($3/$15).

For most of the automation work we do, inbox triage, document review, lead routing, that premium buys you nothing. A cheaper model does the job just as well. Fable 5 earns its keep on long-horizon agentic work, the multi-step builds where a weaker model wanders off and a stronger one simply gets to the end. It still edges Kimi K3 on agentic benchmarks, though nowhere near enough to justify the price gap on its own.

Our take

We reach for Fable 5 when the task is long and the cost of getting it wrong is high. Everywhere else, we use something cheaper. Boring, correct, and it shows up in your margin at the end of the month. And if you're on a paid Claude plan, you might be running Fable 5 for free right now, for as long as that keeps getting extended.

Working out which model belongs in which workflow is exactly what our 30-minute audit is for.