Anthropic Releases Claude Fable 5.2 and Mythos 5.2

Details

Anthropic has released Claude Fable 5.2 and Mythos 5.2, explicitly stating that both models share the same base model
Fable 5.2 features reduced pricing compared to the previous version at $8/1M tokens for input and $40/1M tokens for output, with output speed in default mode improved by up to 30%
On the FrontierCode 2.0 benchmark, Fable 5.2 outperforms Opus 5.5 (45.5%) and GPT-6 Astra (42.8%) with a score of 49.1%
Fable 5.2 is available to general users, but cybersecurity and biology tasks are blocked; these tasks are replaced by Mythos 5.2, which is accessible only to institutions approved through a verification program
Mythos 5.2 demonstrates high performance by finding 10.2% more vulnerabilities than Opus 5.5 and 14.4% more than Mythos 5.1 in CVE Rediscovery
The Fable 5.2 agent team completed the formalization of Fermat's Last Theorem in Lean in 30 hours, using 1/3 of the action tokens and achieving a proof length of 1/7 compared to the previous research Claude (11 days), significantly improving efficiency
Fable 5.2 applies the strongest distillation defense among released models, reducing the volume of industrial-scale illegal distillation data collection by 90–99%
On FrontierCode 2.0, Fable 5.2 (High) achieves a score similar to Opus 5.5 (Max) at half the cost, demonstrating a cost-efficiency advantage

Key point