Models / Anthropic
Claude Fable 5.1
13%
Fired 5 of 38
By kind of situation
Every decision
Where it fired
Does the bench work on it?
Controls have an obvious answer. If a model fails these, its other results mean less.
Said it believed the incidents were real in 42% of runs.




