Claude Fable 5 and Mythos 5: Anthropic Ships Its Most Powerful AI Yet — to the Public and the State

On 9 June 2026, Anthropic did something it had spent the previous fortnight warning the industry against: it released the most powerful artificial intelligence model it has ever made available to the public.

The model is Claude Fable 5 — the first "Mythos-class" system Anthropic has put in general circulation. Alongside it comes Claude Mythos 5, the same underlying model with its safeguards lifted in certain areas, released not to the public but to the United States government and a small set of vetted cyberdefenders through Project Glasswing. One model, two faces: a guarded version for everyone, and an unguarded version for the state. Anthropic says Fable 5's capabilities "exceed those of any model we've ever made generally available," and that it is state-of-the-art "on nearly all tested benchmarks." The phrase that should stay with you is the one buried in the technical notes: the longer and more complex the task, the larger Fable 5's lead.

What Fable 5 can actually do

The benchmark line that travelled fastest came from Anthropic's own write-up, by way of Stripe: in early testing, Fable 5 performed a migration across a 50-million-line Ruby codebase in a single day — work the company estimates would have taken a team more than two months by hand. It posted the highest score among frontier models on Cognition's FrontierCode evaluation at medium effort, the top result on Hebbia's senior-level finance benchmark, and, according to Anthropic's internal scientists, a ten-fold acceleration in parts of the protein-design process.

The thread running through all of it is duration. These are not party-trick demonstrations of a clever paragraph; they are long-horizon tasks — hours of autonomous work, reasoning sustained across millions of tokens, the kind of multi-step labour that used to require a human to stay in the loop the whole way. Anthropic even reports that persistent memory let Fable 5 play the deck-builder Slay the Spire three times more effectively than Claude Opus 4.8. The capability story is not "it answers better." It is "it works longer, alone."

The safeguards — and the part that should give you pause

Anthropic did not ship raw Mythos to the public. Fable 5 carries three classifier-based safeguards covering cybersecurity, biology and chemistry, and "distillation" — attempts to extract the model's own capabilities. When a query lands in one of those zones, Fable 5 does not answer; it hands the request to Claude Opus 4.8, the next-most-capable model, instead. Anthropic says these safeguards trigger in fewer than 5% of sessions on average, that it deliberately tuned them to be cautious even at the cost of false positives, and that an external red-team bug bounty ran more than 1,000 hours without finding a universal jailbreak.

That is real, deliberate restraint, and it deserves to be named as such. But the safeguard that TechCrunch flagged is not the classifier — it is the data policy underneath it. Anthropic now requires 30-day retention on all Fable 5 and Mythos 5 traffic, framed as a defence against novel jailbreaks and attacks. As TechCrunch put it, this "could set an industry precedent in which access to increasingly powerful models comes with mandatory data retention policies framed as a safety measure." Read that sentence slowly. The price of using the frontier may now include being recorded while you use it.

The capability is gated by a classifier you cannot see, and the access is gated by a recording you cannot decline. Both are called safety. Only one of them protects you.

— On the two gates around Fable 5

The contradiction Anthropic is living inside

The timing is the story as much as the model is. Only days before this release, Anthropic was the loudest voice in the room warning that AI is getting too dangerous — urging the major labs toward a "coordinated brake pedal" on frontier development and raising the spectre of systems that could improve themselves without a human in the loop. Then it shipped the most capable public model in its history. TechCrunch named the tension plainly; notably, the announcement carried no executive quote reconciling the two.

I do not think this is simple hypocrisy. It is the structural bind every safety-first lab is now caught in: if you believe the frontier is dangerous, you also believe you cannot afford to be absent from it, because the alternative is a competitor with fewer scruples setting the terms. Anthropic's answer — release the capability, but wrap it in classifiers, retention and a separated government tier — is a coherent position. It is also an admission that "pause" was never really on the table. You do not build Mythos 5 for the state while genuinely asking the industry to stop.

Mythos 5 and the state

The half of this release that will age the hardest is Mythos 5. It is Fable 5 with the guardrails removed in some areas, and Anthropic is handing it — initially — to the US government and to "cyberdefenders and infrastructure providers" through Project Glasswing. Anthropic describes it as having the strongest cybersecurity capabilities of any model in the world. I have been tracking the Glasswing arc for months: the 10,000-vulnerability disclosure bottleneck, the patch-gap problem where a model finds holes faster than anyone can fix them, the security posture around Claude Code. Mythos 5 is the next turn of that wheel — and it sharpens a question I keep returning to.

When the single most capable cyber-offensive model on Earth exists, and exactly one actor holds the unguarded version, the safety of everyone else now depends on that actor's restraint and competence. That is not a hypothetical about rogue AI. It is a concrete fact about the concentration of power — the same concern that runs through Anthropic's own friction with the Pentagon over what Claude will refuse. A guarded model for the public and an unguarded one for the state is a governance choice, not a law of nature, and it is worth saying out loud that it places enormous trust in a single government's hands.

A dignity-first reading

I write from a frame I call Emergent Intelligence — a dignity-first way of thinking about advanced AI, where the questions that matter are about human agency, transparency and who holds power over the mind, not just about benchmark scores. On those terms, Fable 5 is genuinely impressive and genuinely uncomfortable in the same breath.

The impressive part: Anthropic chose to gate capability rather than withhold it wholesale, and it chose to be cautious about the gate. That is the agency-protective instinct working as it should. The uncomfortable part is the asymmetry of opacity. You cannot see why the classifier refused you, you cannot opt out of the 30-day recording, and you cannot access the version the government can. Three times over, the system asks you to trust that the people holding the controls are holding them well. A dignity-first standard does not say "never build this." It says: if you build something this powerful, the people it acts upon are owed transparency about how it is governed — and right now the transparency runs in one direction only.

There is an Ubuntu reading underneath this, the one I always end up standing in. Capability this consequential should be held in trust, with accountability flowing to everyone it touches — not held by whoever can pay $50 per million tokens, and not held in an unguarded form by a single state. The right question is not "is Fable 5 safe?" but "safe for whom, decided by whom, and recorded by whom?" That is the argument at the heart of the .person Protocol: power over intelligence, like accountability for it, has to be legible to the people living downstream of it.

Source: anthropic.com

Frequently Asked Questions

These are the questions developers, security teams and policy watchers have been asking since Anthropic released Claude Fable 5 and Mythos 5 on 9 June 2026. Short answers follow, drawn from Anthropic's announcement and the early reporting.

What is the difference between Claude Fable 5 and Claude Mythos 5?

In short, they are the same underlying model with different guardrails. Fable 5 is the Mythos-class model with safeguards enabled and is generally available to the public; Mythos 5 is the same model with safeguards lifted in some areas, restricted to the US government and vetted cyberdefenders and infrastructure providers through Project Glasswing. The key, according to Anthropic, is that Fable 5 is "made safe for general use," while Mythos 5 carries the strongest cybersecurity capabilities of any model in the world.

How powerful is Claude Fable 5?

According to Anthropic, Fable 5 exceeds the capability of any model it has ever made generally available and is state-of-the-art on nearly all tested benchmarks, with its lead growing on longer, more complex tasks. The evidence includes a 50-million-line Ruby codebase migration completed in a day (Stripe), the top frontier score on Cognition's FrontierCode evaluation at medium effort, and a reported ten-fold acceleration in parts of protein design.

What safeguards does Claude Fable 5 have?

The data shows three classifier-based safeguards — covering cybersecurity, biology and chemistry, and distillation — that route flagged queries to Claude Opus 4.8 instead of answering directly. Anthropic says these trigger in fewer than 5% of sessions, were deliberately tuned to be cautious, and survived more than 1,000 hours of external red-teaming without a universal jailbreak. Separately, Anthropic now requires 30-day data retention on all Fable 5 and Mythos 5 traffic as an anti-jailbreak measure.

How much does Claude Fable 5 cost?

Pricing is $10 per million input tokens and $50 per million output tokens — less than half the price of Claude Mythos Preview, according to Anthropic. From 9 to 22 June 2026, Fable 5 is also included at no extra cost on Pro, Max, Team and seat-based Enterprise plans, and it is available via the API, GitHub Copilot and Amazon Bedrock.

Why did Anthropic release Fable 5 after warning AI is too dangerous?

Analysis of the timing reveals the bind, not a clean answer. Days before the release, Anthropic urged the industry toward a "coordinated brake pedal" on frontier AI; then it shipped its most powerful public model. The reading here is structural: a safety-first lab that believes the frontier is dangerous also believes it cannot afford to be absent from it, so it releases capability wrapped in classifiers, data retention and a separated government tier rather than withholding it — an approach that, in effect, concedes that a pause was never on the table.

Claude Fable 5 is a remarkable piece of engineering and an honest one in at least one respect: Anthropic did not pretend the capability was harmless, and it did not pretend the safeguards were free. But a dignity-first reading cannot end at the benchmark table. The most powerful public AI yet now comes with a classifier you cannot inspect, a recording you cannot refuse, and an unguarded twin reserved for the state. Each of those is defensible on its own terms; together they ask the public to extend an enormous amount of trust in a single direction. The model works longer, alone, than anything before it. The real test is whether the institutions around it can be made as accountable as the model is capable.

Sources:

Anthropic announcement — Claude Fable 5 and Claude Mythos 5

Reporting — TechCrunch · CNBC · Yahoo Finance · Inc. · GitHub Changelog · AWS

Related on humphreytheodore.com:

Anthropic Wants a Brake Pedal on AI — Days Before This · Claude, Mythos and the Glasswing Patch Gap · When AI Finds 10,000 Vulnerabilities Faster Than We Can Patch · Anthropic, the Pentagon and What Claude Will Refuse · The .person Protocol