Anthropic, preparing for its initial public offering (IPO), has unveiled its latest advanced artificial intelligence (AI) models for general users. The models include safety measures to block high-risk queries, while a separate security-verified model will be available only to approved institutions.
On June 9, Anthropic launched two new AI models, 'Claude Fable 5' and 'Claude Mythos 5.' Fable 5 is designed for general users, while Mythos 5 will be offered on a limited basis to verified institutions through security collaboration programs like Project Glasswing.
Both models utilize the same foundational architecture, differing primarily in their safety features. Fable 5 will not respond directly to queries that could be misused in areas such as cybersecurity or biological and chemical threats; instead, it will refer these queries to the previous top model, Claude Opus 4.8, with users being informed of this process.
Queries suspected of attempting unauthorized 'distillation' of competitive AI model functionalities are also included in the restrictions. Anthropic stated that this safety mechanism operates in less than 5% of all sessions on average. While some benign requests may be blocked, the settings are intentionally conservative to ensure a safe launch.
Anthropic classified Fable 5 and Mythos 5 as 'Mythos-level' models, which are rated higher than the existing Opus tier. Mythos 5 achieved a score of 78% in the cybersecurity assessment conducted by ExploitBench, surpassing OpenAI's GPT-5.5, which scored 34%, and Anthropic's Opus 4.8, which scored 40%. The previously released Mythos beta model scored 69%.
In the doctoral-level knowledge assessment known as 'The Last Exam of Humanity,' Mythos 5 scored 59% without using external tools like web searches or calculators, outperforming the Mythos beta's score of 56.8%. In the terminal environment coding assessment, Terminal-Bench 2.1, it scored 88%, exceeding GPT-5.5's score of 83.4%.
In the SWE-Bench Pro, which measures general coding ability, Mythos 5 scored 80.3%, while GPT-5.5 scored 58.6% and Google's Gemini 3.1 Pro scored 54.2%. In the knowledge work assessment GDPval-AA, it achieved a score of 1932, higher than both GPT-5.5 and Gemini 3.1 Pro.
Anthropic will retain user data generated by Fable 5 and Mythos 5 for 30 days, which will be used for detecting new attacks and identifying false positives.
Fable 5 is available starting today. Until June 22, it will be offered to existing paid subscribers at no additional cost. After that date, separate usage credits will be required. Anthropic plans to reintegrate Fable 5 into the general subscription plan once sufficient server capacity is secured.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.