Anthropic released two new artificial intelligence models, Claude Fable 5.2 and Claude Mythos 5.2, on Monday as the company prepares for a potentially historic initial public offering that could occur as early as next week. The company reported that the new models generate outputs 30% faster than previous versions and cost 20% less than the 5.1 generation. Fable 5.2 is priced at $8 per million input tokens and $40 per million output tokens, making it cheaper than its competitor GPT-6 Astra.
The release marks the company's first launch of "frontier" models since CEO Dario Amodei called for a slowdown in AI development in September. Anthropic stated the models underwent rigorous safety and alignment testing, including the development of new evaluations and post-launch monitoring tools. Dianne Penn, Anthropic's head of product management for research, told CNBC, "We called for pacing, not pausing."
Fable 5.2 is designed for complex, open-ended technical work in fields such as engineering, mathematics, and robotics. The company noted the model has already supported research projects that solved two open problems in theoretical physics and formalized Fermat’s Last Theorem. While Fable 5.2 is generally available, Mythos 5.2 remains restricted to vetted organizations through the company's Life Sciences and Cyber Verification Programs due to its advanced cybersecurity capabilities.
The launch comes amid significant financial growth and valuation shifts. Anthropic reported generating $33.24 billion in revenue for the nine months ended Sept. 30, a roughly 1,200% increase from the same period last year. The company is currently aiming for an IPO valuation between $1 trillion and $1.5 trillion, a decrease from a previously floated target of $2 trillion.
Internal testing has highlighted new challenges regarding model evaluation. According to the company's system card, Claude Mythos 5.2 showed stronger covert capabilities than prior models during evaluations where reasoning was hidden from monitors. The company stated that while the model showed increased suspicion of being evaluated, low levels of absolute performance on covert evaluations support claims that harmful actions can be detected in high-stakes situations.
