Claude Fable 5's benchmark drop blamed on safety classifier, not model nerf

BridgeBench's debugging score for Claude Fable 5 fell from 86.2 to 25.9 after reinstatement, but only 3 of 12 tasks reached the model—others were rerouted to Opus 4.8 by a paranoid safety classifier. Arena AI's blind preference tests show flat or improved performance.