AI companies spent years warning us that super-powerful models could eventually become risky. Now, one of them is effectively saying that future has arrived.
You see, Anthropic has announced two new models: Claude Fable 5 and Claude Mythos 5. While Fable 5 is being released publicly, Mythos 5 is mostly staying locked away behind government partnerships and trusted-access programs because of what the company describes as extremely advanced cybersecurity and biology capabilities.
That is a pretty wild sentence to write in 2026.
According to Anthropic, Fable 5 is its most capable publicly available model ever. The company says it outperforms competitors in software engineering, research, vision tasks, and long-context reasoning. Anthropic also claims the model can work autonomously on difficult projects for much longer than previous Claude releases.
Naturally, the company included benchmark charts showing Fable 5 beating basically everything else.
Some of the real-world examples are more interesting than the charts, however. Stripe reportedly used Fable 5 to complete a massive Ruby codebase migration in a single day. Anthropic says the same work would have otherwise taken an engineering team more than two months manually.
If true, that is the sort of thing software developers should probably pay attention to.
Anthropic also highlighted Fable 5’s vision capabilities. Apparently the model can rebuild web apps from screenshots and even beat Pokémon FireRed using nothing but raw gameplay visuals. Earlier Claude models reportedly needed elaborate helper systems just to stumble through similar gaming tasks.
I will admit, the Pokémon example is actually a clever way to demonstrate visual reasoning improvements without resorting to another boring benchmark graph nobody cares about.
But really, the most important part of this announcement is not the performance claims. It is the safety restrictions.
Anthropic openly says Mythos-class models create serious risks in areas like offensive cybersecurity and biology research. Instead of fully exposing those capabilities to the public, Fable 5 uses automated classifiers that silently hand certain requests off to the weaker Claude Opus 4.8 model instead.
That means users may not always be interacting with the flagship model they think they are using.
The company says these fallback protections trigger in less than 5 percent of sessions on average, although Anthropic also admits harmless prompts may occasionally get flagged. Cybersecurity, biology, chemistry, and AI distillation requests are the major areas being filtered right now.
Anthropic claims external red-team researchers spent more than 1,000 hours attempting to jailbreak the system without finding a universal bypass. The company also says Fable 5 refused all tested harmful requests involving cyberattack planning and defense evasion.
Of course, every AI company says its safeguards are strong right up until somebody breaks them on Reddit two days later. We will see how long these protections actually hold up once millions of people start hammering the system.
Meanwhile, Claude Mythos 5 remains restricted to select cybersecurity organizations and researchers through Anthropic’s Project Glasswing initiative. According to the company, Mythos 5 currently has “the strongest cybersecurity capabilities of any model in the world.”
That should probably make everyone at least a little uncomfortable.
Anthropic also made some very aggressive biology claims. The company says Mythos 5 accelerated portions of internal drug-design workflows by roughly ten times and generated molecular biology hypotheses researchers preferred over earlier Opus-class outputs around 80 percent of the time.
One especially bold claim involves genomics research. Anthropic says Mythos 5 autonomously analyzed millions of cells spanning 138 species and built a machine learning model that supposedly outperformed a recent model published in Science while being dramatically smaller.
As always with frontier AI announcements, independent verification matters. AI companies love publishing huge claims before outside researchers have time to validate anything.
Still, even if Anthropic’s marketing numbers are inflated, the broader trend is obvious. Frontier AI companies are no longer pretending these systems are harmless chatbots. We are now entering an era where certain models may only be accessible to approved governments, researchers, and security organizations.
That is absolutely wild.
Anthropic says Claude Fable 5 is available now for $10 per million input tokens and $50 per million output tokens. Mythos 5 access remains heavily restricted for the time being.
Support independent tech journalism
NERDS.xyz is independently owned and operated. If you enjoy my coverage of Linux, AI, hardware, cybersecurity, and tech culture, consider supporting the site on Ko-fi.
Support NERDS.xyz