Anthropic drops Claude Fable 5 with built-in fallback guards

Anthropic just gave the public its first taste of a Mythos-class AI model with the launch of Claude Fable 5, but the safety leash is tighter than ever.

Rolled out on June 9, Fable 5 is billed as Anthropic's most capable publicly available model, setting high-water marks across coding, visual understanding, and complex knowledge work. Alongside it, a less-restricted sibling called Claude Mythos 5 is rolling out to vetted cyberdefenders and infrastructure operators via Project Glasswing in partnership with the US government. However, the debut was far from smooth: Anthropic quickly suspended access to both models due to an unspecified service disruption, promising to restore availability soon.

Brains of a genius, safety filters of a helicopter parent

Fable 5 shares the same core engine as Mythos, but Anthropic wrapped it in heavy-handed safety classifiers. Fearing that frontier capabilities could supercharge cyberattacks or biological weapon development, Anthropic programmed Fable 5 to automatically punt sensitive queries over to Claude Opus 4.8.

While Anthropic claims this handoff occurs in less than 5% of sessions on average, real-world tests show the guardrails can be aggressively overprotective. As reported by The Verge, Fable 5 flatly refused to answer basic high school biology questions, declining prompts about cell membranes, mRNA vaccines, asthma medicine, and what mitochondria do. Anthropic spokesperson Paruul Maheshwary admitted the team chose to be overly conservative on biology to get the model out the door quickly without catastrophic risk.

There is another catch for enterprise teams: Anthropic is enforcing a mandatory 30-day data retention policy on all Fable 5 and Mythos 5 traffic to catch novel jailbreak attempts, even overriding existing zero-data-retention customer agreements.

Crushing benchmarks, when it actually answers

When you keep it far away from biology and exploits, Fable 5's benchmark gains look massive.

Stripe reported that Fable 5 compressed months of work into days, migrating a 50-million-line Ruby codebase in roughly 24 hours. The model also beat Pokémon FireRed using nothing but raw vision inputs, and it scored 90% on data analytics platform Hex's benchmark for complex analytical reasoning. Vibe-coding tool Base44 said it can routinely one-shot complete applications.

In raw science, internal trials showed Mythos 5 speeding up aspects of protein design by roughly 10x, yielding nine viable drug candidates across 14 targets. In genomics, it spent a week autonomously training a tiny machine learning model that outperformed a recently published Science paper.

Both models cost $10 per million input tokens and $50 per million output tokens—double the price of Opus 4.8, though half the price of the previous Mythos Preview. Claude Pro, Max, and Team subscribers can test Fable 5 for free until June 22, after which it will shift to a paid credit system.

Why it matters

If you are an engineer facing a terrifying legacy codebase migration, or a data team dealing with massive financial audit trails, Fable 5 looks like an immediate must-try. The steep token pricing will sting, but as early testers like Rakuten noted, an autonomous model that accurately validates its own work pays for itself if it cuts weeks of senior developer time.

Yet for many organizations, the trade-offs will be tough pills to swallow. The forced 30-day prompt logging kills adoption for privacy-regulated enterprise pipelines right out of the gate. Meanwhile, everyday researchers in healthcare and biology are essentially locked out from the model's actual intellect, left dealing with annoying Opus fallbacks for completely benign homework questions.

We think Anthropic's routing architecture—running a frontier brain behind a strict AI chaperone that downgrades queries on the fly—is a sign of how all lab releases will look moving forward. But releasing a system so terrified of its own capabilities that it refuses to explain cellular biology shows that frontier safety remains an awkward, blunt instrument.