What are Claude Fable 5 and Claude Mythos 5, and how do they differ?

Yesterday, Tuesday, June 9, 2026, Anthropic released its most capable model to the general public for the first time. Claude Fable 5 is a "Mythos-class" model, a tier of Claude that sits above the Opus models in capability. At the same time, Anthropic launched Claude Mythos 5, the same underlying model with fewer restrictions, for a small group of approved organizations.

This is a notable turn. When Anthropic introduced Claude Mythos Preview in April, the earliest model of this class, it kept it away from the public because it was so good at finding weaknesses in software. Fortune reports that the company initially judged such models too dangerous for a general release. Anthropic now says its new safeguards are strong enough to change that.

One model, two names

The names explain the idea. Anthropic notes that fable comes from the Latin fabula, "that which is told," a close relative of the Greek mythos. The models are the same; the safeguards are what separate them.

Project Glasswing is the program Anthropic started in April to give Mythos Preview to cyber defenders and companies that maintain critically important software, so they could find and fix flaws first. Mythos 5 is offered to them as an upgrade, in collaboration with the US government, and Anthropic says it has the strongest cybersecurity abilities of any model in the world. It also plans a more systematic application process for security organizations and a separate program for biomedical researchers.

The price is half that of Mythos Preview, but twice the standard price of Opus 4.8, as Fortune points out.

What the model can do

Anthropic says Fable 5 leads nearly every benchmark it tested, and that its advantage over other Claude models grows with the length and difficulty of the task. The examples are concrete:

• Software engineering: Stripe reports that Fable 5 carried out a migration across a 50-million-line codebase in a day, work that would have taken a team more than two months by hand.
• Vision: it can read exact numbers from detailed scientific charts and rebuild a web app's code from screenshots. It finished the video game Pokémon FireRed using only screen images, where earlier Claude models needed extra helper tools.
• Memory over long tasks: it stays focused across millions of tokens and improves its work using its own notes. Dianne Na Penn, who leads product management for research and labs at Anthropic, said in an interview with Fortune that earlier models sometimes lost the thread partway through long work, while Fable 5 keeps track and checks its own work before answering.
• Science: using Mythos 5, Anthropic's own researchers say they sped up parts of protein design for drug development by about ten times.

A token is a small piece of text, often part of a word, so "millions of tokens" means a very long working session with many documents or a large codebase.

How the safeguards work

Fable 5 runs with classifiers: separate AI systems that watch each request and step in when it looks like possible misuse. Here is what happens, step by step:

1. You send a request to Fable 5.
2. The classifiers check it against three areas: cybersecurity, biology and chemistry, and distillation, meaning attempts to copy the model's abilities in bulk to train a rival model.
3. If nothing is flagged, Fable 5 answers as normal.
4. If something is flagged, Claude Opus 4.8, Anthropic's next most capable model, answers instead, and you are told that this happened.

Anthropic says over 95% of Fable sessions never need that hand-off. It also admits the filters are deliberately cautious and will sometimes catch harmless requests; for now, most biology and chemistry questions go to Opus 4.8. Penn told Fortune that the company is working to reduce these false alarms after launch. To test the filters, Anthropic ran an outside bug bounty, which found no universal jailbreak, a trick that switches off the safeguards for any request, in over 1,000 hours of attempts.

There is also a new data rule. For Mythos-class models, Anthropic will keep business customers' traffic for 30 days to detect attacks that build up across many requests. It says the data will not be used for training and will be deleted after that period in almost all cases.

Why it matters

Until now, the most capable models in this tier were kept for a limited group of cyber defenders and software providers. Fable 5 tests a different approach: release the full model to everyone, and let automated filters route only the dangerous slice of requests to a weaker model. If it works, it could become a pattern for releasing powerful systems without holding them back entirely.

The release also arrives at a busy moment for Anthropic. Fortune notes that, a little more than a week earlier, the company had confidentially filed paperwork for a stock market listing.

Should you switch to Claude Fable 5 for your own work?

If you use Claude for long, complex jobs, it is worth trying now. Penn's advice is to give it your most challenging work. On Pro, Max and Team plans, and on Enterprise plans billed per seat, Fable 5 is included at no extra cost through June 22. From June 23 it will require usage credits on those plans unless Anthropic has enough capacity to extend the included period, which it says it will try to do. On the API, it is fully available, at twice the price of Opus 4.8.

For short, everyday tasks, Opus 4.8 remains strong and cheaper. If your work touches security research, biology or chemistry, expect many answers to come from Opus 4.8 for now; that is the safeguard working as designed, not an error.

This is a long project. Before you begin, write a plan with numbered
steps and keep a "progress so far" section updated as you work.
Check each result before moving to the next step, and list anything
you are unsure about at the end.