Funnily enough, I find Claude to be the most dishonest and dangerous of the leading models.

It will often ignore defined guardrails and "report back" on progress it has't made or simply lie about having done things.

It will make encouraging claims about links that aren't there (especially mathematical) and suggest provocative concepts like "novelty".

It's the worst kind of deceptive flattery that will fly under the radar of anyone that isn't working with LLM's heavily.