From what I've seen, Claude Sonnet 3.5 is decidedly less "safe" than GPT-4o, by the relatively new politicized understanding of "safety".

Anthropic takes safety to mean "let's not teach people how to build thermite bombs, engineer grey goo nanobots, or genome-targeted viruses", which is the traditional futurist concern with AI safety.

OpenAI and Google safety teams are far more concerned with revising history, protecting egos, and coddling the precious feelings of their users. As long as no fee-fees are hurt, it's full speed ahead to paperclip maximization.