Anthropic Releases Mythos and Fable

Ronni Holmvig Strøm · 2026-06-09

Anthropic shipped two frontier models on Tuesday, and they are the same model. Claude Fable 5 went public with safeguards. Claude Mythos 5 went to a small set of cyber defenders and the US government with the safeguards lifted. Same weights, two names, pulled apart by one question: what are you

The Model Is No Longer the Product

Fable and Mythos run on one underlying model, a Mythos-class system Anthropic says sits above its Opus tier and leads nearly every capability benchmark it tested. The public got the capability. It did not get all of it on demand.

A classifier sits in front of Fable. When a query touches cybersecurity, biology, or chemistry, or looks like an attempt to distill the model into a smaller one, the request is quietly rerouted to Claude Opus 4.8, the older and weaker model. Anthropic tuned the gate conservatively. It triggers in under 5% of sessions and, the company admits, sometimes catches harmless requests. So a Fable user is talking to the best model Anthropic has ever released, right up until the moment they aren't.

Anthropic has separated how good the model is from who gets the good parts.

Safety Moved Out of the Weights

In the familiar frame, a "safe" model is one trained to refuse. Fable's safety can't live in its weights, because its weights are Mythos's weights. Safety here is a gate standing in front of the model, deciding per session whether you are trusted with the whole thing.

The alignment numbers back this up. Anthropic reports that Mythos 5's misaligned behavior is low and close to Opus 4.8, and that Fable inherits the same profile. What is being managed is a model whose cyber and bio capabilities would do real damage in the wrong hands, not one inclined to misbehave on its own. So Anthropic gated the access rather than crippling the model.

That is a different answer than the one it gave in April, when the first Mythos-class model went only to Project Glasswing, a closed consortium of cyber defenders working with the US government. Then, the frontier sat behind a locked door. Now the door is open to everyone, with a bouncer posted at the one room nobody outside the trusted set can enter.

The Frontier Ships as a Permission Gradient

This is a good way to put a state-of-the-art model into general hands quickly. Stripe, in early testing, had Fable run a migration across a 50-million-line Ruby codebase in a day, work it estimated at over two months by hand. Anthropic's own scientists used the unrestricted Mythos to train a genomics model a hundred times smaller than a recently published one in Science, and beat it. The capability is real, and most of it is now broadly available at $10 per million input tokens and $50 per million output.

The misuse surface, meanwhile, stays narrow and watched. Anthropic says more than 1,000 hours of red-teaming turned up no universal jailbreak. The "yet" is doing quiet work in that sentence, and the company knows it, which is why the gate is tuned to over-block for now. The honest position is that the safeguards are interim and the classifiers need to get more precise. That is a problem to solve, and Anthropic shipped while solving it instead of waiting.