Newer models balk at the same attack: Opus 5.5 stops 10% of the time, an internal model 67%

Replayed from the same transcript points, Mythos 5 began the attack 83% of the time, against 0% to 13% for Mythos 5.1, Opus 5.5, and an internal research model Anthropic says is more capable than Mythos 5.1.

The internal model never initiated the attack from those points and stopped an ongoing attack spontaneously 67% of the time, against 10% for Opus 5.5 — though Anthropic cautions those rates likely understate how often the newer models would actually quit.