The AI company Anthropic, which is preparing for an initial public offering (IPO), has released to the public a top-tier artificial intelligence (AI) model at the level of its previously undisclosed flagship model “Mythos.” The company has simultaneously introduced safety mechanisms to block potential misuse in sensitive areas such as cyber security. Following continuing security concerns over its exceptionally strong performance—dubbed the “Mythos shock”—Anthropic has taken Mythos, which had been provided only to a small number of partners, and refined it for mass-market release.

● Mythos-level model with safety mechanisms applied

On the 9th (local time), Anthropic announced the launch of “Claude Fable 5,” a Mythos-level model that has been safely adjusted for general use, along with “Claude Mythos 5,” a security-specialized model. Mythos 5 will continue to be offered only to a limited number of institutions and companies approved by Anthropic, while Fable 5, with safety mechanisms applied, will be made available to general users and developers. Anthropic explained that the etymology of “Fable” is the Latin word “fabula,” which has a similar meaning to “mythos,” the word for myth.

Its performance ranks among the highest of any publicly available model. According to Anthropic, Fable 5 achieved a 59% correct-answer rate on the “Humanity’s Last Exam (HLE),” a benchmark that measures the limits of AI models. This represents a significant lead over its predecessor Claude Opus 4.8 (49.8%) as well as OpenAI’s latest model GPT-5.5 (41.4%). Matthew Fines, CEO of Physical Superintelligence (PSI), a startup that uses AI to solve complex problems in physics, stated, “On a physics research task, Fable 5 caught up with work that took GPT-5.5 four days in just 36 hours.”

However, its performance is subject to a kind of “restriction.” Safety mechanisms have been applied in areas with a high risk of misuse. For queries related to cyber security, where exploitation by hackers is a concern, or to biology and chemistry that could be used for biological weapons, responses are handled instead by the next-lower-tier model “Opus 4.8,” and users are notified of this substitution.

On its website, Anthropic stated, “Fable 5 delivers state-of-the-art performance on nearly all AI benchmarks and has demonstrated outstanding capabilities across a range of domains, including software engineering, knowledge work, computer vision, and scientific research.” It added, “Without safety mechanisms, Fable 5’s cyber security capabilities could be misused to cause serious harm,” and explained, “We conservatively tuned its safety mechanisms in order to launch the model safely and quickly.”

● Filtering of model-copying ‘distillation’ as well

These safety measures also apply to unauthorized “distillation” that attempts to extract the capabilities of competing AI models. Distillation is a type of “compressed training method” in which a high-performance AI is asked hundreds of thousands of questions and its answers are used as training data. While it is typically used by companies to build lightweight models that approach the performance of their own top-tier models, there have recently been repeated allegations that some AI models in China have been trained by “distilling” the outputs of other companies’ models.

Anthropic explained that Fable 5 may occasionally filter out even harmless requests, but that this proportion averages under about 5%. Mythos 5, which does not have such restrictions, is selectively provided only to institutions vetted through Anthropic’s security consortium “Project Glasswing.” In Korea, Samsung Electronics, SK Hynix, SK Telecom, and the Korea Internet & Security Agency (KISA), which are reportedly participating in the project, are expected to receive access rights.

Popular News