Claude Fable 5 Performance Drops on Benchmarks, But Safety Classifier—Not Model—Blamed for Routing Failures

According to BridgeBench AI and Arena.AI, Claude Fable 5's reinstatement on July 1 triggered conflicting benchmark results. BridgeBench reported debugging scores collapsed from 86.2 to 25.9, but data showed nine of twelve tasks were rerouted to Opus 4.8 by Anthropic's new safety classifier rather than reaching Fable 5 itself. Meanwhile, Arena.AI's thousands of human-preference votes found Fable 5 performance largely flat or improved across most categories when the model actually handled requests, with document performance up 34 Elo points and expert text up 25.

The distinction matters: general users in creative writing, research, and text analysis will see minimal difference, while developers working on code repair and debugging face constant fallback routing. Anthropic acknowledged the new classifiers cast too wide a net in blocking exploit-related prompts and said refinements will come over time, but provided no timeline.