Back in April, Anthropic introduced its "Mythos" model to the world. Mythos Preview, reportedly, is such a powerful model that it can find security flaws across all kinds of software. In the wrong hands, bad actors could abuse the model to find vulnerabilities in programs, services, and sites most of us rely on for modern digital life. In effect, Mythos could open up the biggest hacking opportunity in history. What a pitch.

As such, Anthropic pulled the brakes on Mythos. While it maintained that it would eventually release the model to the public, it first needed to trial it with a limited pool of trusted testers, in what it calls "Project Glasswing." To start, that meant opening up the model to the U.S. and other governments. While Mythos is still not available to the likes of you or me, Anthropic is releasing a new model that promises many of the capabilities of Mythos, without the accompanying cybersecurity risks.

What are Anthropic's Fable 5 and Mythos 5?

On Tuesday, Anthropic announced its latest model, Claude Fable 5, which it calls a "Mythos-class model" that is "safe for general use." The company says Fable 5 is supposedly better and more capable than any of its other public models. Anthropic claims Fable 5 scores at the top of most benchmarks, including software engineering, knowledge work, vision tasks, and research. The company goes so far as to say "the longer and more complex the task, the larger Fable 5's lead over our other models." There's also Mythos 5, which seems to be Fable 5 without certain limitations, but isn't available to the general public.

According to Anthropic's benchmarking, Fable 5 and Mythos 5 alike outperform Mythos Preview, Opus 4.8, OpenAI's GPT-5.5, and Google's Gemini 3.1 Pro, in the following categories: agentic coding, knowledge work, spatial reasoning, tool use, legal, multidisciplinary reasoning (without tools), biology, cybersecurity, and health. Mythos Preview ekes out a win in computer use and multidisciplinary reasoning (with tools), but it's a clean sweep over all other models.

Anthropic says Fable 5 was able to complete a coding project that would have taken a team over two months to finish in just a day. It can rebuild a web app's source code from only screenshots. It can beat Pokémon FireRed with a "minimal, vision-only harness," while other Claude models struggled to play at all. It was able to play Slay the Spire and reached the final act three times more often than Opus 4.8 Mythos 5 builds on its research abilities, with improved stats in drug design, as well as novel hypotheses regarding questions of molecular biology, and the ability to produce novel research in genomics.

How is Anthropic keeping Fable 5 safe?

That's the big question: If Fable 5 is Mythos-class, how can you ensure that it's safe to release to the general public? Couldn't a bad actor take advantage of Fable 5's capabilities and force it to discover and disclose security vulnerabilities?

Anthropic says it has that figured out. While Fable 5 may be Mythos-level in many ways, the company says that its Project Glasswing testing has produced a model with the proper safeguards for a public release. Fable 5 looks out for "classifiers," or highly sensitive topics, that it knows it should not answer. What that means is this: When Fable 5 receives a request that it thinks has to do with cybersecurity, biology, chemistry, or distillation, it doesn't answer the question itself. Instead, it passes the query off to Opus 4.8, Anthropic's "next-most-capable" model. The model should still be powerful enough to provide accurate answers, but not capable of providing malicious users with the tools necessary to exploit others.

Anthropic says its new guardrails are cautious and conservative, and may be overkill. Benign requests may accidentally trip Fable 5's security alarms, but that supposedly happens around 5% of the time. As such, Anthropic says Fable 5 is able to handle requests itself roughly 95% of the time. In addition, the company found that after a bug bounty program, no white hat hacker could find a universal jailbreak (or an exploit to bypass safety protocols) after 1,000 hours of testing. While one organization has made progress in finding one jailbreak, Anthropic says it's confident that its protocols make it impractical for hackers to discover jailbreaks before the company does.

Why drop requests for biology and chemistry? Anthropic says that Mythos is also too good at aiding gene therapy research and development, which can be a benefit to scientists, but a major risk in the wrong hands. In addition, Anthropic knows that there are actors out there trying to "distill" Claude models' abilities to train their own models to do whatever they want. As such, any of these requests is booted to a lower-performing model.

Anthropic is also making a change to its data retention policy for Fable 5 and Mythos 5. With these models, the company will keep your data for 30 days—not for training, but to help protect against future cyberattacks and jailbreaks. Fable 5 and Mythos 5 are both priced the same: $10 per million input tokens, and $50 per million output tokens, which Anthropic says is less than half the price of Mythos Preview.