Anthropic Unveils Top-Tier Mythos-Class AI Model 'Fable 5' to the Public

Anthropic has released its top-tier AI model, which had previously been withheld from public release, to general users.

Reuters Yonhap News Agency

On June 9 (local time), Anthropic announced the launch of its new AI model, "Claude Fable 5," along with a security-focused model, "Claude Mythos 5." Fable 5 is intended for general users, while Mythos 5 will only be made available to institutions that have passed security verification.

Both models are fundamentally based on the same architecture, but differ in terms of accessibility. Fable 5 incorporates safety mechanisms for areas with high risk of misuse. If a user submits a cybersecurity-related query that could be exploited for malicious hacking, Fable 5 will not answer directly; instead, the previous top-tier model, Opus 4.8, will handle the request. Users are informed when this occurs.

Anthropic explained that these safety measures have been conservatively configured to ensure a safe and rapid rollout. As a result, even legitimate requests may occasionally be blocked, though the company stated that the actual block rate is less than 5 percent of all sessions. These restrictions apply not only to cybersecurity concerns but also to queries related to biological or chemical topics that could be misused for bioweapons. Suspicious requests that appear to be unauthorized attempts to extract features from competing AI models ("distillation") are also subject to these restrictions.

Mythos 5, which has no such restrictions, will be provided only to institutions verified through the security consortium "Project Glasswing." As a result, organizations reportedly participating in this project—including Samsung Electronics, SK hynix, SK Telecom, and the Korea Internet & Security Agency (KISA)—may be granted access to Mythos 5.

Anthropic has also implemented a new policy to retain data generated on the Fable and Mythos models for 30 days. This data will be used to prevent new attacks or to identify false positives.

The performance of the new models surpasses that of the previously released "Mythos Preview." In the "ExploitBench" test, which measures cybersecurity capabilities, Mythos 5 achieved a score of 78 percent, outperforming OpenAI GPT-5.5's 34 percent, Anthropic Opus 4.8's 40 percent, and the Mythos Preview's 69 percent.

In "The Last Test of Humanity," which assesses doctoral-level intelligence across various fields without the use of tools, Mythos 5 scored 59 percent, higher than the previous Mythos Preview's 56.8 percent. In the "Terminal-Bench 2.1" test, which evaluates coding ability in a terminal environment, it scored 88 percent, surpassing GPT-5.5's 83.4 percent.

Fable 5 also achieved high marks for general coding performance, recording 80.3 percent on the "SWE-Bench Pro" benchmark. GPT-5.5 scored 58.6 percent, and Google Gemini 3.1 Pro scored 54.2 percent. In the "GDPval-AA" test, which evaluates knowledge work capabilities, Fable 5 scored 1,932 points, exceeding both GPT-5.5 and Gemini 3.1 Pro.

Fable 5 is available to users starting on June 10, 2026. Until June 22, existing paid subscribers can access it at no additional cost. After that, a separate fee will be required. Anthropic stated that once server capacity is sufficient, Fable 5 will again be included in existing subscription plans.

This content was produced with the assistance of AI translation services.