Anthropic Claude Mythos tops reasoning, security benchmarks

Claude Mythos represents a category-defining leap that moves LLMs from general-purpose assistants to specialized, autonomous reasoning agents. The reported 97% USAMO score confirms a level of mathematical reasoning signaling a major breakthrough in logical consistency, while advanced offensive cyber capabilities identify vulnerabilities with a success rate far exceeding previous models. The introduction of the Capybara tier suggests Anthropic is bifurcating its lineup into consumer and high-compute expert tiers. Restricted access via Project Glasswing highlights growing safety concerns regarding dual-use models, and high operational costs indicate that wide commercial deployment remains distant.