Anthropic Ships Its Most Powerful AI Days After a Risk Warning

Fable 5 puts Anthropic’s frontier Mythos model in public hands, but it refuses high-risk questions and logs every session for 30 days.

Anthropic just handed the public its most powerful AI model for the first time, launching Claude Fable 5 on Tuesday. The twist: it lands days after the company urged AI labs to install a coordinated “brake pedal.” Fable comes with hard guardrails and costs double Opus 4.8, at $10 per million input tokens.

Picture a supercar maker unveiling its fastest model ever, then bolting on a governor that throttles the engine the second you steer into certain neighborhoods. That is roughly the play here. The frontier capability is real and it is now in your hands, ready to one-shot full apps and crank through long analytical tasks. Just do not ask it how to build a bioweapon.

What Happened

Fable 5 is the first publicly available version of Anthropic’s Mythos model. Anthropic says it shines at software engineering, knowledge work, and vision.

The catch is built in. In high-risk areas like cybersecurity, biology, chemistry, and distillation, Fable blocks the response and quietly falls back to Claude Opus 4.8. According to early data, that handoff is rare, with at least 95% of Fable sessions running entirely on the model’s own answers.

Access opens through the Claude API and consumption-based Enterprise plans right away. Subscriptions get a staged rollout. Through June 22, Fable 5 is included in Pro, Max, Team, and seat-based Enterprise plans at no extra cost. On June 23, Anthropic pulls it from those plans and shifts to usage credits, with a promise to restore it as a standard feature as soon as it can.

Anthropic is also pushing a new version, Mythos 5, to organizations already cleared for the advanced model.

The Backstory

Mythos launched as a preview back in April, locked to a handful of partners over cybersecurity concerns. Last week, Anthropic widened the gate to hundreds of organizations across 15 countries, leaning on groups that run critical infrastructure.

The public launch also lands as Anthropic preps for the public markets, alongside OpenAI and Elon Musk’s SpaceX. And it follows the company’s own warning that AI systems are advancing so fast they may soon hit recursive self-improvement, the point where models improve themselves with no human in the loop.

So the message is layered. Anthropic is the lab waving the caution flag and the lab shipping the fastest engine. Fable 5 is how it tries to hold both ideas at once.

The Plan

The safety story is the product story. Anthropic says it stress-tested Fable’s classifiers with jailbreak attempts before release.

An internal bug bounty ran over 1,000 hours and produced no universal jailbreaks. External red-teaming orgs came up empty too.

Then there is the data move. With Fable 5 and Mythos 5, Anthropic now requires 30-day retention on all traffic, even for enterprises that previously had zero-retention deals. The company says it will not train on that data and will only use it to defend against novel attacks and to cut false positives. That single policy could become an industry template, where access to the most powerful models comes bundled with mandatory retention framed as safety.

The early reviews are strong. Analytics firm Hex said Fable was the first model to hit 90% on its benchmark of complex, long-running analytical tasks, praising its judgement on the hardest questions. Vibe-coding platform Base44 flagged better “one-shotting” of full apps and excellent tool-calling. Workspace platform Genspark said Fable beat every other model it tested, with standout UI design and game coding.

The Business Model Angle

Here is the pattern worth bookmarking: Anthropic is selling the constraint, not hiding it. The guardrails, the red-teaming receipts, the retention policy are all packaged as features for buyers who answer to compliance teams. Safety is the differentiator, and it is being priced at a premium.

And it is a real premium. Fable 5 runs $10 per million input tokens and $50 per million output tokens, exactly double Opus 4.8. Anthropic is betting that the top of the market will pay for a model that, in Rakuten’s words, reflects on and validates its own work, where the extra thinking pays for itself.

Watch the rollout choreography too. Preview for partners, then critical infrastructure, then the public, then a deliberate price wall on June 23. That is a scarcity ladder. Each rung builds demand and trains buyers to treat access as a privilege rather than a default. For founders, the lesson is that how you release can matter as much as what you release.

The Risk

The honest counterpoint starts with the bill. Double the price of an already pricey model is a deterrent, full stop. Many enterprises are already turning sour on AI costs after watching the invoices land or blowing through annual budgets early, and frontier models that split one request into many tasks make that worse.

The fallback cuts both ways too. Even if deferral to Opus 4.8 is rare, “rare” is not “never,” and an inconsistent answer quality is a hard thing to sell at a premium.

Then the retention reversal. Telling zero-retention customers they now get 30-day retention, no opt-out, may read as safety to some and as a broken promise to others. Privacy-sensitive buyers are exactly the high-value accounts a frontier model needs.

Anthropic itself admits demand will be very high and difficult to predict. Translation: this could strain capacity, pricing, and the goodwill the whole “responsible lab” brand runs on.

Quick Questions

The Bottom Line

Anthropic is trying to be the lab that warns about the fire and sells the fastest match, and Fable 5 is the test of whether buyers will pay a premium for caution baked in. For founders, the takeaway is sharper than the AI drama: your constraints can be your pitch, your release schedule can manufacture demand, and the right buyers will pay extra for trust. The open question is whether double the price and a rewritten privacy promise are a moat or a ceiling.