Claude Fable 5 tops RadLE 2.0 — why 'I don't know' beats accuracy
radle 2.0 benchmark shows frontier models hallucinate with high confidence; claude fable 5 leads by prioritizing safety over raw accuracy.
crash lab introduced radle 2.0 to evaluate if ai models can safely handle radiology cases by recognizing when to defer to humans. testing 16 models across 200 cases, the benchmark penalizes high-confidence errors and rewards uncertainty, revealing that human radiologists still outperform ai on the combined accuracy-confidence metric. claude fable 5 topped the primary metric by safely refusing difficult cases, while gemini 3 pro had the highest raw hit rate. medical-specific and open-weight models performed poorly by attempting to answer every question with dangerous confidence. this signals that while model accuracy is approaching human levels, their inability to self-assess remains a major barrier for clinical deployment.