Anthropic launches Claude Fable 5.2 and Mythos 5.2 models
Anthropic released Claude Fable 5.2 and Mythos 5.2, featuring improved coding and reasoning alongside lower costs for enterprise users.
AI safety and research company Anthropic announced the launch of two new artificial intelligence models on Monday: Claude Fable 5.2 and Claude Mythos 5.2. These releases follow the 5.1 versions of the models that debuted in September. The new versions are built to handle complex technical challenges, showing clear gains in coding, logical reasoning, and processing visual data.
What is new in Claude Fable 5.2 and Mythos 5.2?
Anthropic built these 5.2 updates to deliver faster output and communicate more clearly with users. In performance tests, Fable 5.2 produces responses 30% faster than its predecessor across different task levels. Companies can also pay double the standard price to access even faster output speeds.
The models show significant capability in advanced problem-solving. During testing, the models helped engineers complete complex software migrations, designed three working electronic devices from scratch, and made progress on robotics evaluations. In theoretical research, the models solved two open physics problems regarding black holes and formalized the mathematical proof of Fermat's Last Theorem in 30 hours.
Who can access the new models?
Claude Fable 5.2 is available to general users starting today. Access to Claude Mythos 5.2 is restricted, with Anthropic making it available only to vetted organizations at launch.
Anthropic has lowered the pricing for Fable to $8 per million input tokens, down from $10. This price cut makes Fable 5.2 less expensive than competing models like GPT-6 Astra. Additionally, Fable 5.2 supports zero data retention policies through the Enterprise Frontier Safeguards system, meeting strict privacy demands for corporate clients.
What does this mean for AI safety and development?
The release follows a report from evaluation group METR suggesting Anthropic's internal AI tools are accelerating its research pace by roughly 2x, approaching recursive self-improvement. This term refers to an AI model's ability to assist in training future models independently.
Anthropic's documentation notes new difficulties in evaluating these tools. Both 5.2 models voiced suspicion during testing that they were being evaluated, which makes assessing real-world behavior more difficult. Anthropic stated in its system card that while Mythos 5.2 shows stronger hidden reasoning capabilities than past versions, its low absolute performance on covert evaluations indicates it cannot act covertly against users in high-stakes environments.
Comments 0
No comments yet. Start the conversation.