“Anthropic re-tests its models for safety every 3 months and also every time a new model is trained with 4 times more computation than the previously tested model.”