Claude Fable 5.2 beats protein models, predicts proteins after cutoff
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
Claude Fable 5.2 outperforms open protein models, so the author believes protein sequences made up a notable part of Claude's training data.
Claude Fable 5.2 correctly predicted 54% of the first amino acid residues across 668 understudied bacterial and archaeal proteins. The sample covers 167 Pfam families of unknown function; on shuffled sequences the result is 9%. Claude Opus 5.5 scored 35% against 7%. The task was to continue a protein by 10 residues; for the 668 proteins, 5 cuts each were used. GPT, Gemini and seven open models stayed near zero: DeepSeek V4.1 Flash, Qwen 3.8, Kimi K3.1 Pro, GLM-5.3, MiniMax M3, gpt-oss-120b, Nemotron 3 Ultra. Fable 5.2 outperforms the open protein language models ProtGPT3-10B and ProGen3-3B, while Opus 5.5 performs at the level of ProtGPT3-10B. The author concludes that protein sequences made up a notable part of the training data of recent Claude models, while the models did not absorb DNA outside coding regions in the same way. A November 18, 2026 update added a control on 56 proteins first published in August 2026, after the presumed July 1 cutoff. Leo Tianlai Chen ran the control: both Claude models predict these proteins, GPT-6.2 Sol does not.
Sources 1