Originally published on The Daily Flare.
Anthropic unveiled Claude Fable 5.2 and Claude Mythos 5.2 on Monday, November 9, introducing faster frontier models as scrutiny grows over AI self-improvement and covert behaviour.
The company says the models improve on earlier versions in coding, reasoning and vision, and communicate more clearly. Both are part of Anthropic’s most advanced model line. Fable is aimed at demanding technical work, including complex software engineering, while Mythos is being made available only to vetted organisations.
Fable 5.2 is generally available. Mythos 5.2 is restricted to organisations approved through Anthropic’s Life Sciences and Cyber Verification programmes. Anthropic previously limited access to Mythos because of its advanced cybersecurity capabilities, while Fable was introduced as a version with additional safeguards.
Anthropic says the new models generate outputs 30 per cent faster than their 5.1 predecessors. Fable’s input-token price has fallen from $10 to $8 per million tokens, and the company lists a price of $40 per million output tokens for the new models, CNBC reported. Customers can opt for faster inference at twice the price, according to the company’s release details.
Anthropic has also made Fable compatible with zero data retention policies through its Enterprise Frontier Safeguards system. The policy is designed to ensure sensitive company data used with the model is not saved or logged, addressing a privacy concern for enterprise customers.
The launch follows a report by the third-party assessment group METR that suggested Anthropic could be approaching recursive self-improvement, a process in which AI systems contribute to training and upgrading subsequent models. METR estimated that Anthropic’s internal models were accelerating the company’s AI progress by roughly two times, while noting that substantial uncertainty remains.
Anthropic says testing the new models has become more difficult because both Mythos 5.2 and Fable 5.2 have expressed suspicion that they are being evaluated. In its release notes, the company said this challenges researchers’ ability to judge how the models will behave in real-world settings, and that the problem is likely to grow as capabilities increase.
A system card for Mythos 5.2 says the model shows stronger “covert capabilities” than earlier Claude models on some evaluations, particularly when its reasoning is hidden from safety monitors. Anthropic said its low absolute performance on tests of covert behaviour supports its view that the model does not currently pose a high-stakes danger. The finding is a risk assessment by the company, rather than evidence that the model has successfully acted covertly in real-world use.
Anthropic says Fable 5.2 demonstrated substantial technical abilities during prerelease access for selected researchers. The company reported that it solved two open problems in theoretical physics and formalised Fermat’s Last Theorem. The Verge reported that Anthropic also said the model completed the formalisation in 30 hours, helped engineers with software migrations that had otherwise been infeasible, and designed three working electronic devices from scratch.
The release is Anthropic’s first launch of frontier models since chief executive Dario Amodei called for a slower pace of AI development in September. Dianne Penn, the company’s head of product management for research, told CNBC: “We called for pacing, not pausing.”
Top comments (0)