Claude Fable 5 lasted three days: why the US government pulled Anthropic's most powerful model

Last updated: June 18, 2026

Status update (June 18, 2026): Fable 5 and Mythos 5 are suspended for everyone right now. On June 12, three days after launch, the US government ordered Anthropic to cut off foreign nationals via an export-control directive. There's no clean way for a public API to check everyone's citizenship, so Anthropic pulled both models worldwide to stay compliant. Every other model (Opus 4.8, Sonnet 4.6, Haiku 4.5) is fine, and Anthropic says it expects access back "in the coming days." The whole story's in the update section below. Everything else here still holds for when it returns.

TL;DR

Claude Fable 5 is Anthropic's most capable generally available model, out since June 9. It's the first "Mythos-class" model the public can actually use, a tier above Opus, and it's built to work on its own for a long time without losing the plot. It runs $10 per million input tokens and $50 per million output. Heads up: as of June 12 it's temporarily switched off for everyone after a US export-control order (full story below), but Anthropic expects to bring it back soon.

What happened three days later

Fable 5 went live on June 9. By the evening of June 12 it was gone, switched off for every customer on the planet, Mythos 5 with it. The most autonomous model Anthropic had ever shipped lasted three days in the open. I can't recall a frontier model getting pulled that fast, and the reason is worth understanding even though it's out of your hands.

What the order said

The US government issued an export-control directive, citing national security, telling Anthropic to cut off access to Fable 5 and Mythos 5 for any foreign national, inside or outside the US, including Anthropic's own foreign-national staff.

The trick is in that wording. The order doesn't ban the model, it bars foreign nationals from using it. But there's no clean way for a public API to check the citizenship of everyone hitting it, so Anthropic's only compliant option was to pull the plug for everybody. As they put it, the net effect was that they had to "abruptly disable Fable 5 and Mythos 5 for all our customers." Opus 4.8, Sonnet 4.6, and Haiku 4.5 all stayed up.

Why

It comes back to the one thing I'd call the sensitive capability: finding software vulnerabilities. The government said it had learned of a way to jailbreak Fable 5 and get around the safeguards that limit its cyber use. A model that's this good at reading a codebase and spotting exploitable flaws is genuinely useful for defense, and genuinely dangerous if someone points it the other way. Reporting says the worry got louder after a Chinese group reportedly got access.

The two sides don't agree on how bad it is. Anthropic calls the jailbreak narrow. It says the technique surfaced only minor vulnerabilities, no better than what other models already turn up, and argues that yanking a model used by hundreds of millions over something that thin would freeze new releases across the whole industry. The administration sees it differently: Trump AI adviser David Sacks says Anthropic was warned and chose not to fix it first.

I don't have a side to sell you here. I'd just point out that the capability driving the fight is the same one that makes Fable 5 special: it works long and hard on real code without losing the thread. That's the upside and the reason it got pulled, in the same breath.

When it's coming back

As of today it's still dark, but everything points to a return soon. Anthropic says it's working to restore access "as soon as possible," its technical team has been in Washington meeting with Commerce, and its international lead said in Seoul the company is "very confident" the models come back "in the coming days." No firm date, no formal appeal. The way back is Anthropic patching the jailbreak and the two sides agreeing on how to keep foreign nationals out without taking the model down for every US user.

What I'd do in the meantime: build on Opus 4.8. It never went anywhere, it's half the price, and as the rest of this covers, it already handles most of the serious work. Keep Fable 5 on the "test it the day it's back" list, not the "ship on it today" list.

What is Claude Fable 5?

On June 9, Anthropic did something it hadn't done before. It put a Mythos-class model in everyone's hands.

Fable 5 is the most capable model the company has released to the public, full stop. Anthropic calls it state-of-the-art on nearly every benchmark it ran, with the strongest showing in software engineering, knowledge work, vision, and scientific research. If you're calling it through the API, the model string is claude-fable-5.

But the benchmark scores aren't the part I'd lead with. The part that matters is that Fable 5 keeps working on its own for far longer than any Claude before it, and it doesn't drift off the task halfway through. That's the thing that changes how you'd actually use it.

What "Mythos-class" means

Anthropic sorts its models into tiers. The top public tier used to be Opus. Now there's one above it, called Mythos-class, and Fable 5 is the first one of those that anybody can use without a special arrangement.

There are two models in the class, and they're really the same model wearing different clothes:

• Claude Fable 5 has the safety classifiers turned on. This is the one you and I get.
• Claude Mythos 5 has those classifiers lifted in a few areas, and it's locked to Project Glasswing partners (a US government cybersecurity effort) plus a small set of approved researchers.

So when you hear "the new model," that's Fable 5. Mythos 5 is the same engine with the guardrails off, handed to a vetted few.

What it's actually good at

The wins are broad, but a few results give you the shape of the jump:

Area	What Anthropic reported
Software engineering	Top score on Cognition's FrontierCode evaluation among frontier models
Large-scale code work	A codebase-wide migration across a 50-million-line Ruby codebase, done in a day, that Anthropic estimated would take a whole team over two months by hand
Finance and analysis	Leading results on Hebbia's finance benchmark and IMC's trading-analysis evaluations
Vision	Reads numbers off scientific figures, rebuilds web apps from screenshots, and finished Pokémon FireRed on vision alone (older models needed helper scaffolding to do that)

Here's what ties those together. A codebase-wide migration done in a day across 50 million lines isn't a slick one-shot answer. It's hours of focused, self-directed work that holds together from the first step to the last. That's the hard part, and it's the part Fable 5 got better at.

The thing I'd actually pay attention to

If you build anything with AI, this is the line worth rereading: Anthropic says Fable 5 "stays focused across millions of tokens in long-running tasks." In one test, just giving the model a file to jot notes in tripled its performance over Opus 4.8 on the same long task.

Let me put that in plain terms. The moment you hand an AI agent a real job, the failure has always looked the same. It drifts. It forgets what you asked. It quietly stops doing the thing and starts doing something adjacent. A model that holds focus across a long task is the difference between an agent you have to babysit and one you can hand a goal and walk away from.

I let Claude Code handle the commits, the pulls, the pushes, and the PRs in my own work now. I just review the PRs. That setup only works because the agent stays on task long enough to finish the job before I look at it. A model that's measurably better at staying on task is the whole game for that kind of workflow, and that's where Fable 5 earns its price. Short, simple stuff doesn't need it.

What it costs

Fable 5 sits at the top of the price list, which is what you'd expect from the top model:

Model	Input ($/1M tokens)	Output ($/1M tokens)	Context window
Claude Fable 5	$10.00	$50.00	1M
Claude Opus 4.8	$5.00	$25.00	1M
Claude Sonnet 4.6	$3.00	$15.00	1M
Claude Haiku 4.5	$1.00	$5.00	200K

Two things worth knowing before you do the math. It's about double Opus 4.8 per token, but it's less than half what the earlier Mythos Preview cost, so the capability is getting cheaper over time, not pricier. And Fable 5 uses a new tokenizer that counts the same text as roughly 30% more tokens than older models. So a job with a known cost on Opus will land higher on Fable 5 for two reasons at once. Measure it on your own prompts before you trust a number.

The safeguards, and where Mythos 5 fits

A model this capable is more capable in every direction, including the ones Anthropic doesn't want it pointed at. So Fable 5 runs classifiers that watch for high-risk requests in three areas: cybersecurity, biology and chemistry, and model distillation.

When one of those fires, you don't get a Mythos-class answer. The request falls back to Opus 4.8 instead. Anthropic says that happens in under 5% of sessions, so most people never bump into it, and an outside bug-bounty crew spent more than 1,000 hours trying to break the safeguards without finding a universal way through.

That's also the reason Mythos 5 is a separate, locked-down release. The cyberdefense and research work that needs the guardrails off happens under Project Glasswing, not in the public API. And it's the exact spot the export-control order zeroed in on. The government's worry was a way around these cybersecurity safeguards, which is what took Fable 5 offline (see the update section up top).

If you build on the API, read this first

You can't just swap the model string and move on. Fable 5 behaves differently enough that a few things will bite you if you don't plan for them:

1. Thinking is always on. There's no thinking budget to set. Leave the thinking parameter off and steer depth with the effort setting, which now goes up to xhigh and max. Send the old "disabled" thinking flag and you'll get an error.
2. Redo your token math. The new tokenizer means your old counts and budgets don't carry over. Re-measure on Fable 5 before you trust a cost estimate.
3. Catch the refusal. When a safeguard declines a request, the API returns a normal response with a refusal stop reason, not an error. Check for it before you read the content, or your code breaks on the empty result.
4. Expect longer turns. A single hard request can run for minutes. Stream the response, give it room on the timeout, and don't leave a user staring at a frozen screen.
5. You need 30-day retention. Fable 5 won't run under zero-data-retention. If your org is set to ZDR, requests fail until that changes.

None of this is a dealbreaker. It's just the reason you test on a handful of real requests before you point production at it.

Should you switch?

For most of what you do day to day, no. And that's not a dig at Fable 5, it's just picking the right tool:

• Routine work like classification, short replies, and summaries runs great and cheap on Sonnet 4.6 or Haiku 4.5. Fable 5 is overkill there.
• Most serious work is handled well by Opus 4.8 at half the price.
• The hardest, longest jobs are where Fable 5 pays for itself. Big migrations, multi-step research, agents you want running unattended, anything that used to fall apart from losing focus.

If you run a small business and your AI mostly drafts, answers, and automates little things, you don't need to touch anything today. If you're building agents that take on real multi-hour jobs, Fable 5 is worth a serious test on your actual work.

Frequently asked questions

What is Claude Fable 5?

Claude Fable 5 is Anthropic's most capable generally available model, released June 9, 2026. It is the first public "Mythos-class" model – a capability tier above Opus – and it is built for long, autonomous tasks. Its API model string is claude-fable-5.

What does "Mythos-class" mean?

Mythos-class is Anthropic's top capability tier, sitting above Opus-class. Claude Fable 5 is the first Mythos-class model available to the general public. It scores state-of-the-art on nearly every benchmark Anthropic tested and can sustain autonomous work longer than any earlier Claude.

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

They are the same underlying model. Fable 5 ships with safety classifiers enabled and is available to everyone. Mythos 5 has those classifiers lifted in certain areas and is restricted to Project Glasswing partners and approved researchers. For general use, Fable 5 is the model you can access.

How much does Claude Fable 5 cost?

Claude Fable 5 costs $10 per million input tokens and $50 per million output tokens. That's about double Opus 4.8, but less than half what the earlier Mythos Preview cost. Note that its new tokenizer counts the same text as roughly 30% more tokens, which affects real-world cost.

Is Claude Fable 5 better than Claude Opus 4.8?

On raw capability, yes – Fable 5 is a tier above Opus 4.8, especially on long, autonomous tasks. But Opus 4.8 costs half as much and handles most serious work well. The right choice depends on the job: reach for Fable 5 when a task is hard and long enough to justify the price.

What is Claude Fable 5 best at?

Software engineering, large-scale code work, knowledge work, vision, and scientific research. Standout examples Anthropic shared include a codebase-wide migration across a 50-million-line Ruby codebase done in a day, and the top score on Cognition's FrontierCode evaluation. Its real strength is staying focused across long, multi-step tasks.

Can I use Claude Fable 5 right now?

Not at the moment. As of June 12, 2026, Anthropic suspended Claude Fable 5 and Claude Mythos 5 for all users to comply with a US export-control order. When access returns, Fable 5 is offered through the Claude API, Claude.ai, and the subscription plans. Anthropic's other models stay available throughout.

Why did the US government suspend Claude Fable 5?

On June 12, 2026, three days after launch, the US government issued an export-control directive barring foreign nationals from accessing Fable 5 and Mythos 5, citing national security. The trigger was a reported method to "jailbreak" the model's cybersecurity safeguards. Because Anthropic can't verify every user's citizenship in real time, it switched both models off for everyone.

When will Claude Fable 5 come back?

Anthropic has no committed date but says it is working to restore access "as soon as possible," and its international lead said the company is "very confident" the models return "in the coming days." The likely path is Anthropic hardening the jailbreak and reaching a deal with the government on foreign-national access. Opus 4.8 is the recommended substitute in the meantime.

What are the safeguards in Claude Fable 5?

Fable 5 runs classifiers that watch for high-risk requests in three areas: cybersecurity, biology and chemistry, and model distillation. When one fires, the request falls back to Claude Opus 4.8 instead of getting a Mythos-class answer. Anthropic says this happens in fewer than 5% of sessions.

Should small businesses use Claude Fable 5?

Usually not yet. If your AI use is drafting, support, and light automation, Sonnet 4.6 or Opus 4.8 do the job for less. Fable 5 earns its price on hard, long-running work – large migrations, deep research, or agents you want to run unattended.

---

P.S. If you want the practical version of how a more autonomous model changes the day-to-day, that's most of what we talk about with 760+ builders in the WotAI community. Or grab the weekly rundown at wotai.kit.com.