앤트로픽의 파격적인 승부수: '클로드 페이블5'가 여는 AI의 새로운 지평 > K-wave Trends

The newly released Claude Fable 5 shares essentially the same computational foundation as Anthropic's elite model, Mythos 5. However, Fable 5, provided to general users, is equipped with robust safety mechanisms to defend against sensitive areas such as cybersecurity, biological risks, and unauthorized model distillation. When a specific high-risk query is entered, the AI makes an immediate judgment and switches the response entity to the smaller Opus 4.8 model, transparently notifying the user of this process. This conservative safety design is interpreted as Anthropic's sophisticated strategic choice to achieve both rapid market entry and social responsibility. Although it operates in less than 5% of total sessions, this filtering clearly demonstrates the intent to prevent AI from becoming a weapon in the hands of malicious hackers.

In terms of technical performance, Fable 5 and Mythos 5 shatter all existing benchmark records, demonstrating overwhelming dominance. On the 'ExploitBench,' which measures cybersecurity capabilities, Mythos 5 recorded a phenomenal score of 78%, easily surpassing the competitor's latest model, GPT-5.5. Furthermore, it achieved a 59% accuracy rate on the 'Human Last Exam (HLE),' which evaluates PhD-level intelligence, marking the first time a model has crossed the impregnable 50% barrier. It showed performance that outclassed rival models in coding tasks and knowledge-based work efficiency, proving that Fable 5 has evolved from a simple language model into a practical agent capable of solving complex logical tasks on its own. These figures are indicators that Anthropic has reached the pinnacle of AI, not just in speed, but in practical problem-solving ability.

It is noteworthy that Mythos 5 is provided selectively only to strictly vetted institutions through the security consortium 'Project Glasswing.' Leading domestic companies and organizations such as Samsung Electronics, SK Hynix, SK Telecom, and the Korea Internet & Security Agency (KISA) are participating in this consortium, and attention is focused on how they will integrate this top-tier AI capability into industrial sites. Anthropic has established an advanced data policy that retains data generated from these models for 30 days to learn about new cyberattacks and identify false positives. This is a glimpse into Anthropic's ecosystem strategy to build a security infrastructure that continuously evolves within a closed environment, going beyond simple model distribution. This approach of building a strong defense system with trusted partners is highly likely to become the standard for the enterprise AI model market.

The pricing structure introduced with the launch of Fable 5 clearly shows how revenue models in the AI market are changing. While it is provided to existing paid subscribers at no extra cost until the 22nd, a pay-as-you-go system—effectively a premium pricing model for top-tier models—has been introduced thereafter. Although the price of $10 per 1 million input tokens and $50 per output token is the highest among existing API services, the prevailing view is that it is sufficiently competitive for enterprises, considering the efficiency of handling complex tasks with fewer tokens. Anthropic has stated plans to include it in subscription plans once server capacity is secured, but for the time being, the highest-performance model will maintain this high-cost 'pay-as-you-go' strategy. This is interpreted as an intention to maximize its value as a premium service while signaling to the market that AI computational resources are a precious commodity.

Field test results show that Fable 5's autonomous task performance differs significantly from existing models. As seen in the case of fintech company Stripe, the ability to complete the migration of a massive codebase of 50 million lines in just one day has the potential to boost human developer productivity by dozens of times. Furthermore, its ability to process visual information in real-time to independently clear complex games suggests that AI has entered a stage where it understands and acts within physical and logical environments, beyond simply processing text. The robust safety mechanisms, which passed over 1,000 hours of jailbreak attempt testing, also provide a basis for the safe adoption of this model in corporate environments. This practical performance will be the strongest growth driver Anthropic presents to investors ahead of its IPO.