Frontier models struggle to classify genetic variants as harmless or harmful

Lewis Tunstall says he tested three frontier models on classifying 1,000 genetic variants as harmless or harmful, and none gave a usable result across the set.

GPT-6.1 Sol returned bare labels with its reasoning encrypted. Gemini 4 Argon refused more than 50% of prompts, and Opus 5.5 ran out of tokens on about 20%. Tunstall adds that open-weight models performed barely better than random.