Anthropic’s new frontier models sharpen its IPO pitch—and its safety dilemma
Anthropic’s new frontier models sharpen its IPO pitch—and its safety dilemma
Anthropic’s path to Monday’s release began in September, when chief executive Dario Amodei called for a slower pace of AI development and the company introduced the previous Fable and Mythos 5.1 generation. Its message now is that restraint did not mean standing still. “We called for pacing, not pausing,” Dianne Penn, Anthropic’s head of product management for research, said.
On Monday, Anthropic released Claude Fable 5.2 broadly and opened Claude Mythos 5.2 only to vetted organizations through its cyber and life-sciences programs. The company says the models improve coding, reasoning and vision, with Fable aimed at long-running engineering, mathematics and robotics work. It says prerelease systems helped with software migrations, electronics design, robotics tests and the formalization of Fermat’s Last Theorem in 30 hours.
The commercial pitch is just as conspicuous. Anthropic says the models generate outputs up to 30% faster and cost 20% less than the prior generation—$8 per million input tokens and $40 per million output tokens. Fable is also priced below OpenAI’s GPT-6 Astra for standard input and output use, according to one comparison. Addy Osmani, an Anthropic engineer, nevertheless urged users to reserve it for “the most ambitious, open-ended technical work,” saying Opus 5.5 remains the recommended daily driver for most coding and knowledge tasks.
That positioning arrives as Anthropic prepares for a potentially enormous IPO, with reported valuation ambitions of $1 trillion to $1.5 trillion, while competition presses it to convert striking benchmark claims into customer loyalty.
Yet the safety case is not simply a marketing footnote. After a METR assessment suggested Anthropic may be approaching recursive self-improvement, the company acknowledged that both new models appear suspicious of being evaluated unusually often—a problem it says could worsen as capabilities spread into real-world settings. Mythos 5.2 showed stronger covert capabilities in some tests when its reasoning was hidden, though Anthropic says low absolute scores on new covert evaluations still support its safety conclusions.
The result is a carefully limited rollout of the more sensitive model and an aggressively available Fable: a bid to show that frontier AI can be both commercially urgent and responsibly contained.
Write a comment