First time going through a model’s system card; wanted to note interesting stuff I found
System Card Link: https://cdn.sanity.io/files/4zrzovbb/website/037f06850df7fbe871e206dad004c3db5fd50340.pdf
Model training
No details on this at all, basic yap about how they trained on internet data respected people blah blah + worked with ethical companies
Single-turn evaluations
Opus 4.7 and Mythos use “adaptive thinking” mode
TLDR: Anthropic cheaping out on the reasoning efforts of their models to save a buck