Claude Fable 5 vs Mythos 5: Same Model, Different Safety

On June 9, 2026, Anthropic released two models on the same day: Claude Fable 5 and Claude Mythos 5. The names look like two different products. They aren't. Underneath, they're the same Mythos-class model. The thing that separates them is safety, not capability, and that one difference changes who's allowed to use which.

Quick disclosure first. ShadowProof is independent and we're not affiliated with Anthropic. We make a US mobile IP app for creators, and we write about the tools our readers keep asking about. This is just us reading the release and laying it out plainly. If you only remember one line from this piece, make it this: almost everyone will only ever touch Fable 5, and for normal coding, writing, analysis, and creative work, none of this safety machinery shows up.

Same model, two safety settings

Start with the part people get wrong. Fable 5 and Mythos 5 are not a "small" and a "big" version. They share the exact same underlying model. Same weights, same training, same raw capability. If you handed both the identical prompt for a marketing email or a Python script, you'd get the same quality of answer.

What's different is how locked down each one is.

Fable 5 is the public release, and it's the careful one. It ships with safety classifiers sitting in front of the model. When a request lands in one of three sensitive areas, the classifier catches it and hands that request back to Opus 4.8 instead of letting Fable 5 answer. So the public model has a few doors that are quietly bolted shut.

Mythos 5 is the same model with some of those bolts removed. It's the unguarded version. And because it's unguarded, Anthropic doesn't hand it out to the public at all. More on who does get it in a minute.

Why split one model in two

The reasoning Anthropic gave is blunt. Frontier models are getting capable enough that, in a few specific domains, the same skill that helps a defender could help an attacker. Cyber and bio are the obvious ones. Rather than ship that raw capability to everyone, Anthropic gated those areas for the public model while still letting trusted, vetted people use the version without the guards. One model, two doors, depending on who's knocking.

What the guarded areas are

There are exactly three categories the Fable 5 classifiers watch for. When a request fits one, Fable 5 doesn't answer it directly; Opus 4.8 takes over.

• Cybersecurity. Offensive security work, the kind of thing that sits close to building attacks rather than defending against them.
• Biology and chemistry. The bio and chem space where detailed help could edge toward genuinely dangerous territory.
• Distillation. This one's about protecting the model itself. It means using Fable 5's outputs to train a rival model, basically extracting its capabilities to build a competitor. Anthropic guards against that too.

Here's the number that puts it in perspective: those classifiers trigger in under 5% of sessions on average. Across the wide world of everyday prompts, the guarded zone is a thin sliver. If you write code, draft documents, summarize research, build a landing page, or brainstorm content ideas, you will most likely never set one of these off. And if you do, the result isn't an error. Opus 4.8 just answers instead, and you may not even notice the handoff happened.

Who actually gets to use Mythos 5

Mythos 5 is not on the API. It's not in your Claude plan. You can't buy it. Anthropic restricts it to vetted partners, and it built the model specifically so that defenders and researchers can do legitimate work that the public model blocks.

Right now access runs through two channels:

• Project Glasswing, a cybersecurity program. This is for the defensive side, the people whose job is hardening systems and who need a model that won't refuse the security questions Fable 5 hands off.
• A trusted access program for biology researchers that's opening up. Worth a careful note here: even in this program, the cyber safeguards stay switched on. So "Mythos 5" doesn't mean every guard is gone for everyone. Biology researchers get the bio doors opened, but the cyber doors stay locked for them.

That's the whole point of the design. Lift exactly the safeguards a given trusted group needs, and no more. A vetted bio lab doesn't automatically get the cyber capabilities, and vice versa.

If you're reading this and wondering whether you qualify, the honest answer for almost everyone is no, and that's by design, not an oversight. These are narrow programs for vetted institutions, not a waitlist you join.

What this means for a normal user

Let's bring it back to the practical question: does any of this matter for you?

For the overwhelming majority of people, the answer is no, and that's genuinely good news. You'll use Fable 5. It's the public model, it's the full Mythos-class capability, and the safeguards almost never come up in normal work. The under-5% figure isn't marketing softness, it's just how rarely everyday prompts wander into cyber, bio, or model-distillation territory.

A few honest takeaways:

• You are not getting a weaker model. Fable 5 is the same model as Mythos 5. The capability is identical. The only thing you give up is direct access to three narrow domains, and even there you get Opus 4.8's answer rather than a flat refusal.
• The pricing is identical. Both run at $10 per million input tokens and $50 per million output tokens. There's no premium for the unguarded version, because access isn't something money buys here. It's gated by vetting, not by price tier.
• The data terms are the same too. Both carry a 30 day data retention policy, and Anthropic doesn't train on the data. That holds whether you're a public Fable 5 user or a vetted Mythos 5 partner.

We dig into the public model on its own in our Claude Fable 5 explainer, and if you want the full rundown of how the handoff classifiers behave, our guide to Fable 5's safety guardrails goes deeper on exactly when they fire.

So the "vs" in "Fable 5 vs Mythos 5" is a little misleading. You're not picking between them. One is the public model you'll actually use, and the other is a restricted variant of that same model for a small set of trusted defenders and researchers. Knowing both exist is useful. Choosing between them isn't really a choice most people will ever face.

Frequently asked questions

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

They're the same underlying Mythos-class model. The difference is safety scope. Fable 5 is the public, locked-down version with safety classifiers that hand cybersecurity, biology and chemistry, and distillation requests back to Opus 4.8. Mythos 5 is the same model with some of those safeguards lifted, and it's restricted to vetted partners rather than released publicly.

Can I use Mythos 5?

For almost everyone, no. Mythos 5 isn't public. It's limited to vetted partners through two channels: a cybersecurity program called Project Glasswing and a trusted access program opening for biology researchers. If you're a normal API user or on a Claude plan, you're using Fable 5, not Mythos 5.

Is Fable 5 less capable than Mythos 5?

No. They're the same model, so the raw capability is identical. Fable 5 just has three narrow domains gated by safety classifiers, and in those cases it hands the request to Opus 4.8 rather than answering. Those classifiers trigger in under 5% of sessions on average, so for everyday coding, writing, and analysis you get Fable 5's full capability.

Why did Anthropic release two versions?

Anthropic warned that frontier models are getting powerful enough to be genuinely dangerous in areas like cyber and bio. So it gated those areas for the public release (Fable 5) while still letting trusted defenders and researchers use the unguarded version (Mythos 5) to do legitimate work the public model blocks. One model, two safety scopes, matched to who's using it. You can see how this compares against the older model in our Fable 5 vs Opus 4.8 piece.