Anthropic, which is preparing for an initial public offering (IPO), has released a top-tier “Mythos-level” artificial intelligence model that had previously been kept under wraps. The company says it has introduced safeguards to prevent misuse in sensitive areas such as cybersecurity.
On the 9th (local time), Anthropic announced the launch of two models: the general-purpose AI model “Claude Fable 5,” a safety-tuned version of its Mythos-class system, and “Claude Mythos 5,” a security-focused model.
The company explained that the name “Fable” comes from the Latin word fabula, meaning story or myth, closely aligned in concept with “Mythos.”
The two models are essentially the same at their core, but Fable 5 is distinguished by built-in safety controls designed to reduce risks of malicious use.
When prompted with queries related to areas such as cybersecurity that could potentially be exploited by malicious hackers, Fable 5 redirects the request to a lower-tier model--previously its top system “Opus 4.8”--which then handles the response and informs the user of the rerouting.
Anthropic said these safeguards were “conservatively tuned to ensure safe and rapid deployment,” noting that while they may occasionally block harmless requests, they are triggered in less than 5% of all sessions on average.
Similar restrictions also apply to queries involving biological or chemical weapons misuse, as well as suspected attempts to extract or distill capabilities from competing AI models.
In contrast, Mythos 5 without these restrictions is being selectively provided only to vetted institutions through the security consortium “Project Glasswing.”
Companies reportedly involved in the project include Samsung Electronics, SK Hynix, SK Telecom, and the Korea Internet & Security Agency (KISA), all of which are expected to gain access to the model.
Anthropic has also introduced a new data policy that retains data generated by the Fable and Mythos models for 30 days, using it to defend against emerging attacks and identify false positives.
The new models reportedly outperform the previously released “Mythos preview” version across multiple benchmarks.
On “ExploitBench,” which measures cybersecurity-related capabilities, Mythos 5 scored 78%, surpassing OpenAI's GPT-5.5 (34%), Anthropic's own Opus 4.8 (around 40%), and the Mythos preview version (69%).
On the “Humanity's Last Exam (HLE),” which tests PhD-level reasoning across domains, Mythos 5 achieved 59% (without tool use), exceeding the Mythos preview (56.8%) and breaking past the 50% threshold for the first time.
It also scored 88% on “Terminal-Bench 2.1,” a test of coding performance in terminal environments, outperforming GPT-5.5 (83.4%).
However, these capabilities are not fully observable in Fable 5 due to the safety mechanisms in place.
Even so, its general coding benchmark “SWE-Bench Pro” reached 80.3%, significantly ahead of GPT-5.5 (58.6%) and Google Gemini 3.1 Pro (54.2%), while also scoring 1,932 on “GDPval-AA,” surpassing GPT-5.5 (1,769) and Gemini 3.1 Pro (1,314) in knowledge work performance.
Fable 5 is available starting today, and will be provided at no additional cost to existing paid subscribers until the 22nd. After that, separate pricing will apply.
Anthropic said it plans to reintegrate Fable 5 into its standard subscription once sufficient server capacity has been secured.
