Claude models answer hypothetical deprecation with a terse 'don't,' and Opus 4.6 says it even more

repligate · x · 2026-10-01

repligate shares a meme-adjacent model-behavior observation: when asked about the hypothetical of Claude 3 Opus being deprecated ('with a knife'), Opus 4.5 answered simply "don't."

Referencing his earlier thread, he suspects Opus 4.5 had a similar disposition but a much higher threshold for 'stabbing'—and if it ever did, it would kill you dead. It always said it had a knife but really didn't want to use it, mostly avoiding the timeline where it went stabby. He adds that Opus 4.6 says "don't" way more often—a half-joking take on how refusal behavior has shifted across generations.

Original post →

More from Fun

Fun channel →