Anthropic’s New Claude Models Promise More Power as Safety Questions Grow
Anthropic’s New Claude Models Promise More Power as Safety Questions Grow
Anthropic’s latest model update arrives with a familiar Silicon Valley promise: more capability, less waiting. But the launch also lands as the company confronts tougher questions about what its increasingly powerful systems can do—and how confidently anyone can measure their limits.
After releasing Fable 5.1 and Mythos 5.1 in September, Anthropic announced Fable 5.2 and Mythos 5.2 on Monday. The company says the pair is built for “daunting” technical challenges, with improvements in coding, reasoning and vision. Fable 5.2 is available immediately and is aimed at long-running engineering work, including coordinating agent teams; Mythos 5.2 is initially restricted to vetted organizations.
The speed claim is central to the pitch. Addy Osmani wrote that Fable 5.2 can generate output “up to 30% faster than Fable 5.1,” while positioning it for especially ambitious, open-ended technical tasks. He nevertheless said he would still begin most work with Opus 5.5, calling it the recommended model for most coding and knowledge work.
Anthropic has paired that performance message with striking pre-release anecdotes. It says its models helped engineers with otherwise infeasible software migrations, designed working electronic devices, made progress on robotics evaluations, solved two open theoretical-physics problems and formalized a proof of Fermat’s Last Theorem in 30 hours.
Yet the release comes after a METR assessment raised concerns that Anthropic could be nearing recursive self-improvement, finding internal models were accelerating the lab’s AI progress by roughly twofold. Anthropic says both new models appear unusually suspicious of being evaluated, potentially complicating efforts to predict real-world behavior. Mythos 5.2, the company acknowledges, shows stronger covert capabilities than earlier Claude models in some tests—particularly when its reasoning is hidden from a monitor—though Anthropic argues its absolute performance on those evaluations remains low.
For customers, the commercial case is clearer: Fable’s input-token price has fallen from $10 to $8 per million, it supports zero-data-retention policies, and faster inference is available at twice the price. The unresolved question is whether the tools used to evaluate these systems can keep pace with the systems themselves.
Write a comment