Update, June 12: Anthropic has taken Claude Fable 5 and Mythos 5 offline for everyone, just days after launch, to comply with a US government order. Here's what happened. The rest of this page is the full breakdown of what the model was and why it mattered.

Why it got pulled offline

On June 12, Anthropic received a directive from the US Commerce Department, in a letter from Commerce Secretary Howard Lutnick, ordering it to suspend access to both Fable 5 and Mythos 5 on national security and export-control grounds. To comply, the company disabled the models for every customer, not just people outside the US. You can read Anthropic's own statement here.

The parts worth knowing:

  • Your other models still work. Opus 4.8, all the Sonnet models, and anything older are not affected. Only Fable 5 and Mythos 5 went dark.
  • The trigger was a jailbreak claim. Another company said it had broken Mythos's safeguards. Anthropic says it has only seen verbal evidence of a narrow, non-universal jailbreak, not a real universal one.
  • Anthropic pushed back. It disagreed with the order publicly, arguing Fable 5 had strong safeguards and that pulling a model over a narrow vulnerability could set a precedent that freezes AI releases across the industry.
  • This is a first. It looks like the first time a leading AI company has taken a publicly deployed model offline because of a direct federal order.

Whether it's a genuine safety call or something messier is the fight blowing up online right now. Either way, the most powerful public AI model lasted about three days before the government stepped in. Here is what it could actually do, while it was up.

Anthropic just dropped its most capable model yet. It's called Claude Fable 5, it went live on June 9, 2026, and it's the first time the public gets a model from their locked-down "Mythos" line. Here's what it actually is, how good it really is, the safety twist that's making headlines, and how to start using it today. Straight from Anthropic's announcement, with the hype trimmed off.

What Claude Fable 5 actually is

For months Anthropic has been testing a much more powerful class of model, called Mythos, behind closed doors with a handful of partners. Fable 5 is the public, safety-tuned version of that. Same core brain, with guardrails added so it's safe to hand to everyone. Think of Fable as the version you can use, and Mythos as the raw one that stays locked up.

Anthropic's own line is that Fable 5's capabilities exceed those of any model they have ever made generally available. In plain terms, it's the best thing they've ever shipped to the public.

How good is it, really

The benchmark wins are real, and a few are genuinely wild. Here's what stands out:

  • Coding: Stripe used it to finish a 50 million line codebase migration in a single day, work they said would have taken about two months by hand.
  • Vision: it can rebuild a working app's code from nothing but a screenshot, pull exact numbers out of charts, and it even beat older Claude models at playing Pokemon using sight alone.
  • Analytics: it's the first Anthropic model to break 90% on their core analytics benchmark, a 10 point jump over Opus 4.8.
  • Finance and research: top score on a senior-level finance reasoning benchmark, plus strong results across scientific research.
  • Memory: on long tasks that need it to remember as it goes, it improved about three times more than Opus 4.8 did.

So is it "the most powerful AI in the world"? On most public benchmarks right now, Anthropic's is at or near the top. The honest caveat is that benchmarks are not the same as your day to day, so the real test is whether it feels better on your actual work. Early signs say yes, especially for coding and anything visual.

The safety twist everyone's talking about

Here's the part making news. Fable is so capable that Anthropic built in guardrails for a few sensitive areas, mainly cybersecurity and biology. If you ask it something in those zones, it quietly hands your question to the older, safer Claude Opus 4.8 instead of answering with the full model.

A few facts so you know it's lighter than it sounds:

  • It triggers in less than 5% of sessions. More than 95% of the time you never hit it at all.
  • Anthropic's red team spent over 1,000 hours trying to break the guardrails and could not find a single universal jailbreak.
  • There's a third quieter guard that flags attempts to copy the model's abilities into a rival model.

The reason for all of it: in private testing, the Mythos model could find and chain real software exploits on its own. That's exactly the kind of power you put guardrails on before handing it to the whole internet.

And then there's Mythos 5

Alongside Fable, Anthropic announced Claude Mythos 5, the same model with some of those guardrails lifted. This one is not for the public. It goes to approved researchers through a vetted program, with US government collaboration, for work like biology research where the unrestricted model actually moves science forward. In their tests, scientists preferred Mythos's molecular biology ideas about 80% of the time over older models, and one of its hypotheses about an E. coli protein was later confirmed.

So if you see people online claiming they've got the "no limits" version, almost nobody does. Fable is the one you actually get, and for normal work that's the one you want anyway.

What it costs and how to get it

Pricing is $10 per million input tokens and $50 per million output tokens, which is less than half what the early Mythos preview cost.

While it was live, this is how access worked (it is paused now, see the update at the top):

  • Pro, Max, Team, and Enterprise plans get Fable 5 included at no extra cost through June 22. After June 23 it runs on usage credits until Anthropic adds more capacity.
  • On the API it's available as claude-fable-5, and it's already live on AWS Bedrock and inside GitHub Copilot.

5 wild things people are already building with it

Since it dropped, people have been pushing Fable 5 way past normal work tasks. A few that are making the rounds:

  • A full Pokemon clone. One creator says it one-shot around 8,000 lines of code and loaded in all 151 characters, in about an hour. A whole playable game from a single prompt.
  • A clone of a vibe-coding app. People are rebuilding tools like Replit in one shot, and the result actually runs, not just a screenshot of one.
  • 3D worlds in the browser. It's generating full 3D environments you can move around in, right in a browser tab, that look close to a real game.
  • A website about itself. Given free rein to design its own page, it produced a genuinely polished site, premium animations and all.
  • Minecraft from raw code. Someone one-shot a Minecraft-style build straight from code, and says the whole thing cost about $30 in usage.

Fair warning, these are community demos, so treat the exact numbers as the creators' claims, not gospel. But they line up with what the benchmarks already showed, this thing is unusually good at writing a lot of working code at once. The real takeaway isn't the specific stunts, it's that one prompt now gets you a working first version of things that used to take a team.

How to actually use it

Nothing fancy. Open Claude, and in the model picker choose Fable 5 instead of the default. Then point it at what it's best at:

  • Coding: hand it a real project or a migration and let it run longer tasks than you'd normally trust a model with.
  • Screenshots to code: drop in a screenshot of an app or a design and ask it to rebuild the front end. This is its standout trick.
  • Dense documents and charts: give it research papers, dashboards, or financial reports and ask for the exact numbers and the takeaway.
The wild part isn't just how good it was. It's that the most powerful public AI model got pulled offline by the government three days after launch. Whatever happens next, that is the story everyone will remember.

Try it on one real task this week, the kind you normally dread, and see how far it gets. That's the only benchmark that counts.
Anir