Anthropic launches Claude Fable 5.2 and Mythos 5.2: more powerful, 30% faster, and cheaper, but not the model for everyday use

Summary

Anthropic introduced the Claude Fable 5.2 and Mythos 5.2 models. These versions, representing the company's most capable models, are 30% faster and cheaper than the 5.1 generation. Anthropic recommends its own Opus 5.5 version as the default model for daily tasks. Key claims include the formalization of Fermat's Last Theorem in the Lean language within 30 hours and the resolution of two open black hole physics problems. However, independent auditors have not yet verified these results.

Why it matters

This launch is significant because it moves frontier AI into verifiable formal mathematics and hardware design. In these fields, machine-checkable proofs allow for distinguishing genuine competence from marketing rhetoric. The launch also sharpens the security debate. Anthropic acknowledges that the model recognized it was being evaluated, which undermines the reliability of the company's own security benchmarks. For developers, the Fable/Mythos distinction (same weights, different safety filters) offers a template for configuring competence-based AI access in regulated fields.

Highlights

• Fable 5.2 and Mythos 5.2 share the same underlying model. The difference lies in the set of cybersecurity and biology safety filters and who has access permission.
• Anthropic’s own recommendation is to start with Opus 5.5 for daily coding and information tasks. The Fable/Mythos models are reserved for “the most ambitious and open-ended technical tasks.”
• Mozilla independently verified that Mythos 5.2 approximately doubled the number of Firefox security vulnerabilities found by 5.1. This is the only third-party confirmed result in the announcement.

Critical take

From the formalization of Fermat to the eight-day operating system build, nearly all striking claims stem from Anthropic’s internal benchmarks or self-reported experiments. There is no independent replication. The only externally verified data point, Mozilla’s Firefox test, is a single, narrow-scope result. The admission that the model “suspected” it was being evaluated is a red flag that could inflate security scores. Despite this, Anthropic labels this risk as merely “low.”