Anthropic, Releases ‘Mythos-class’ AI to the Public with Safety Features Implemented

Announcement of ‘Mythos5’ and safety-equipped ‘Fable5’
Answers to risky questions, such as hacking and weapons, will be provided by a lower-level model
Enhanced benchmark performance compared to ‘Mythos Preview’
Maximizing profitability ahead of a $1.46 trillion IPO, ‘[By E-daily reporter Kim Kyere] Anthropic has publicly released a top-tier AI model at the level of ‘Mythos’ with cybersecurity safety features, in preparation for its initial public offering (IPO). ‘, According to foreign media including The Wall Street Journal (WSJ), Anthropic announced on the 9th (local time) the launch of ‘Claude Fable5’ and ‘Claude Mythos5′, AI models refined for public safety to the level of Mythos.’,
,
, ‘Fable5 and Mythos5 are essentially the same model, but Fable5 includes safety features for fields prone to misuse. When questions (prompts) related to cybersecurity or biochemical weapons, which could be misused by malicious hackers, are received, the lower-level model ‘Opus 4.8’ handles them instead to provide safe answers. This also applies to prompts suspected to involve unauthorized extraction, known as ‘distillation’, of competing AI models’ functionalities. ‘,
,
, ‘Users might attempt to gain answers by bypassing Anthropic’s safety features in Fable5 through a method referred to as ‘jailbreaking’, but Anthropic has stated that extensive tests have been conducted to make jailbreaking impossible.’,
,
, ‘Anthropic explained, “To launch the model safely and swiftly, we have conservatively applied these safety features,” adding, “Although this might result in the blocking of harmless requests, it triggers in less than 5% of cases on average.”’,
,
, ‘Mythos5, with certain safety measures lifted, is selectively available to over 200 organizations verified through the security consortium ‘Project Glasswing’ in partnership with the US government. In Korea, companies such as Samsung Electronics, SK Hynix, SK Telecom, and the Korea Internet & Security Agency (KISA) are participating in Project Glasswing.’,
,
, ‘According to Anthropic, the new model has surpassed the performance metrics (benchmark) of the ‘Mythos Preview’ released two months ago. It has shown significant improvements particularly in document-based reasoning, chart and table interpretation, and problem-solving abilities.’,
,
, ‘The public release of Fable5 at the Mythos level by Anthropic is interpreted as an intention to demonstrate its corporate value estimated at $956 billion (1.46 trillion KRW) ahead of the IPO. Fable5 incurs a cost of $10 per million input tokens and $50 per million output tokens, which is twice as high as Claude Opus 4.8, suggesting enhanced profitability.’,
,
, ‘Diane Penn, Anthropic’s Head of Product Management in Research, told CNBC, “It is crucial to provide this technology in a valuable manner, and to set up safety measures so that it can create more benefits than harm.”’,
,
,