Anthropic Says Its Latest Model Is “worthy of the Mythos Name,” but With Stringent Safety Measures.
In April, Anthropic unveiled its “Mythos” model . The Mythos Preview is reportedly so powerful that it can detect security vulnerabilities in a wide range of programs. In the wrong hands, attackers could use this model to find vulnerabilities in the programs, services, and websites that most of us rely on in our modern digital lives. Essentially, Mythos could open the biggest attack surface in history. What a presentation!
Thus, Anthropic suspended development of Mythos. Although it had stated that it would eventually release the model to the public, it first needed to test it with a limited number of trusted testers as part of the so-called “Project Glasswing.” This initially meant opening the model to the US and other governments. While Mythos remains inaccessible to the likes of you and me, Anthropic is releasing a new model that promises many of Mythos’s capabilities without the accompanying cybersecurity risks.
What are the fifth fable and the fifth mythological romance of Anthropos?
On Tuesday, Anthropic announced its newest model, the Claude Fable 5 , which it calls a “Mythos-class model” and “safe for general use.” The company claims the Fable 5 is supposedly better and more functional than any of its other publicly available models. Anthropic claims the Fable 5 leads in most tests, including software development, intelligence tasks, computer vision tasks, and research. The company even claims that “the longer and more complex the task, the greater the advantage the Fable 5 has over our other models.” There’s also a Mythos 5, which appears to be the same Fable 5 without certain limitations, but it’s not available to the public.
According to Anthropic’s testing results, Fable 5 and Mythos 5 outperform Mythos Preview, Opus 4.8, OpenAI GPT-5.5, and Google Gemini 3.1 Pro in the following categories: agent programming, intellectual labor, spatial reasoning, tool use, legal issues, interdisciplinary reasoning (without tools), biology, cybersecurity, and healthcare. Mythos Preview slightly outperforms all other models in computer use and interdisciplinary reasoning (with tools).
Anthropic claims Fable 5 completed a programming project that would have taken the team over two months in just one day. She can reconstruct the source code of a web app using only screenshots. She can complete Pokémon FireRed with “minimal, vision-based equipment,” while other Claude models struggled. She completed Slay the Spire and reached the final act three times more often than Opus 4.8. Mythos 5 enhances her research capabilities, improving her drug discovery capabilities, generating new hypotheses in molecular biology, and enabling new genomics research.
How does Anthropic ensure Fable 5’s security?
The main question: if Fable 5 is a Mythos, how can we guarantee its safe release to the public? Could an attacker exploit Fable 5’s capabilities and force it to discover and disclose security vulnerabilities?
Anthropic claims to have addressed this issue. While Fable 5 is comparable to Mythos in many ways, the company says testing with Project Glasswing has produced a model with adequate security measures for public release. Fable 5 monitors “classifiers,” or highly sensitive topics known not to be answered. This means that when Fable 5 receives a query it believes is related to cybersecurity, biology, chemistry, or distillation, it doesn’t respond to the query itself. Instead, it passes the query to Opus 4.8, Anthropic’s “next-best” model. This model is designed to be powerful enough to provide accurate answers, but not capable of providing attackers with the tools needed to exploit others.
Anthropic claims its new security mechanisms are cautious and conservative, and perhaps overkill. Harmless requests can accidentally trigger Fable 5’s security alerts, but this is believed to occur approximately 5% of the time. Therefore, Anthropic claims Fable 5 is capable of handling requests independently approximately 95% of the time. Furthermore, the company found that, following its vulnerability bounty program, not a single white-hat hacker was able to find a universal jailbreak (or exploit to bypass security protocols) after 1,000 hours of testing. While one organization has made progress in discovering one jailbreak, Anthropic is confident that its protocols make it impractical for hackers to discover jailbreaks before the company does.
Why are requests for models in biology and chemistry being rejected? Anthropic claims that Mythos is also too good at supporting gene therapy research and development, which can be beneficial for scientists but poses a serious risk in the wrong hands. Furthermore, Anthropic is aware that there are profiteers trying to “usurp” the capabilities of Claude’s models to train their own models to do whatever they please. Therefore, any such requests are redirected to a less effective model.
Anthropic is also changing its data retention policy for Fable 5 and Mythos 5. Under these models, the company will retain your data for 30 days—not for training, but to protect against future cyberattacks and hacks. Pricing for Fable 5 and Mythos 5 is the same: $10 per million input tokens and $50 per million output tokens, which Anthropic says is less than half the price of the Mythos Preview.