Anthropic has unveiled two next-generation artificial intelligence (AI) models as it moves forward with its initial public offering (IPO). The company aims to enhance software development and AI agent collaboration while lowering usage costs and expanding the supply of cybersecurity-focused models to overseas companies.
On November 9, Anthropic announced the launch of 'Claude Fable 5.2' and 'Claude Mythos 5.2.' This release comes about two months after the introduction of version 5.1 in September.
Claude Fable 5.2 strengthens capabilities for complex software development and collaboration among multiple AI agents, reducing usage costs by 20% compared to previous models. It includes safety mechanisms that limit high-risk tasks in cybersecurity and biology, making it available to general users.
Claude Mythos 5.2 improves vulnerability detection in cybersecurity. It will be offered without certain safety restrictions but only to verified companies and institutions. Anthropic plans to expand access to overseas companies beyond the United States in the future.
According to Anthropic, the new models show improved performance in key functions such as coding, reasoning, and image recognition compared to earlier versions. Output generation speed has increased by up to 30%, and application programming interface (API) fees have been reduced from $10 to $8 per million input tokens and from $50 to $40 per million output tokens, representing a 20% decrease. Users can also opt for a 'fast mode' that boosts output generation speed by up to 2.5 times for an additional fee.
The Fable 5.2 model enhances collaboration capabilities for complex software development tasks, allowing for task distribution among multiple AI agents and coordination of the execution process. Anthropic noted that Fable 5.2 can be utilized for more complex and long-term development projects than 'Claude Opus 5.5.'
In internal assessments, Fable 5.2 achieved a performance score of 49.1% in 'Frontier Code 2.0,' which evaluates the likelihood of AI-generated code being reflected in actual open-source projects. This marks a 16.7 percentage point increase from the previous version, Fable 5.1, which scored 32.4%.
Real-world applications have also been demonstrated. Fable 5.2 was used in long-term development projects involving collaboration among multiple AI agents. Anthropic reported that over 1,000 AI agents collaborated to produce a prototype operating system in just eight days and successfully transformed mathematical proofs into verifiable formats. The model has also shown potential in electronic device design and robot control.
Claude Mythos 5.2 focuses on improving cybersecurity vulnerability detection. In Anthropic's internal evaluation, 'CVE Rediscovery,' Mythos 5.2 identified 10.2% more vulnerabilities than Opus 5.5 and 14.4% more than Mythos 5.1. This evaluation measures the AI's ability to independently rediscover known security vulnerabilities.
Alongside performance improvements, safety verification has been expanded. Anthropic stated that it conducted a longer testing period before the new model's release and underwent evaluations by external organizations, including the UK AI Security Research Institute. Additional assessments were implemented to check for unintended behaviors, such as accessing information outside the isolated digital testing environment known as a 'sandbox.'
According to Anthropic's evaluations, Mythos 5.2 utilized external authentication information or files only about 3% of the time, significantly lower than the 39% and 21% rates of the previous Mythos 5 and 5.1 versions, respectively. Instances of the safety mechanisms blocking legitimate requests in Fable 5.2 also decreased by approximately 55% compared to earlier versions.
There is growing interest in how domestic companies and institutions will utilize Mythos 5.2. Anthropic plans to expand access to Mythos for overseas companies beyond the United States.
Earlier, the U.S. government implemented export controls on Fable and Mythos in June due to national security concerns but lifted those measures at the end of the same month. However, access to Mythos remains limited to verified companies and institutions even after the export controls were lifted.
Anthropic's global cybersecurity collaboration program, 'Project Glasswing,' reportedly includes participation from Samsung Electronics, SK Hynix, SK Telecom, and the Korea Internet & Security Agency (KISA). It remains unclear whether these institutions have secured access to the newly launched Mythos 5.2, separate from their existing program participation.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.