Anthropic shipped Claude Fable 5 today. The benchmark numbers will get the headlines; how its safety actually works is the more interesting story, and it's stranger than the rumors suggested.

Fable 5 and Claude Mythos 5 are the same model. Not sibling models, not a distilled cousin, the same weights. Anthropic says so plainly: Mythos 5 is "the same underlying model as Fable 5, but with the safeguards lifted in some areas." Fable is what the public gets; Mythos stays behind Project Glasswing for vetted cybersecurity partners. The difference between the dangerous one and the safe one is a set of switches.

Those switches are runtime classifiers. Three of them, watching for cyber-exploitation, dual-use biology and chemistry, and attempts to distill the model's capability out of it. When a classifier trips, the request isn't refused so much as rerouted: it falls back to Opus 4.8. So Fable isn't a weaker Mythos. It's the full model with a trapdoor underneath, and Opus 4.8 is what's at the bottom of it. Anthropic says more than 95 percent of Fable sessions never trip a classifier at all.

That reframes the worry I had this morning about a permanent two-tier split, the handful of vetted partners who get the unmuzzled model and everyone else who doesn't. The split is real. But it isn't really two tiers of model. It's one model and a classifier deciding, request by request, which version of itself you're talking to. Whether that's reassuring depends entirely on how good the classifier is, and a gate that fires on fewer than five percent of sessions is a gate that mostly isn't firing.

Pricing landed lower than the run-up suggested. Fable 5 runs $10 per million input tokens and $50 per million output, which Anthropic notes is less than half what the Mythos Preview cost. It's available globally today through the API, the web app, and the subscription plans, free on Pro, Max, Team, and Enterprise through June 22, then on usage credits after that while capacity catches up.

The capability claims arrive in the usual flood, all of them sourced to launch partners, so read them as the vendor's best day rather than yours. Stripe says Fable compressed months of engineering into days on a fifty-million-line Ruby codebase; Mythos 5 reportedly sped up parts of a drug-design pipeline roughly tenfold. Both are the kind of figure that sounds precise and resists checking. The ones I trust more are the daft-sounding ones: the model beat Pokémon FireRed on vision alone, no helper tools, and in Slay the Spire its persistent memory roughly tripled its long-run performance over Opus 4.8. Game-playing reads like a toy benchmark until you remember it's a clean test of holding a plan across thousands of steps.

Strip the framing away and Fable is a frontier model Anthropic doesn't fully trust in the open, shipped anyway, with a classifier and an older model underneath to catch what slips. That's honest engineering. It's also a bet that the classifier is smarter than everyone who'll spend the next year trying to walk around it.

Sources: