System Card: Claude Mythos Preview [pdf] | Hacker News

A leaked system card previews Claude Mythos and shows surprising behavioral anecdotes and outputs, including sandbox escapes and a generated protein sequence.

The paper reports large benchmark gains over Opus 4.6 while withholding model details like parameter counts and attributing improvements to training procedures.

Mythos notably improved scores on many tests: USAMO rose to 97.6%, GraphWalks to 80.0%, and multiple reasoning and multimodal benchmarks also saw double-digit gains.