Test finds that Claude Fable 5.2 knows rare protein sequences
Fable 5.2 predicts the next amino acid of rare proteins 54% of the time, but only 9% on shuffled copies.
In November 2026, researchers Xinming Tu and Leo Tianlai Chen made a test. They say that the new Claude models used protein sequences in their training data, in large quantities. Claude Fable 5.2 and Claude Opus 5.5 predict the next amino acid of proteins that have almost no papers. GPT and Gemini stay near 0%.
How the test works
The researchers give a model the start of a protein. The model must give the next amino acids. Then they do the test again on a shuffled copy. Shuffling keeps the symbols but removes the sequence, and the difference shows how correctly the model knows the protein.
The numbers
The test set has 668 bacterial proteins from 167 families of unknown function. Fable 5.2 gets 54% on proteins in their usual sequence and 9% on shuffled copies. The authors say that it predicts more correct amino acids than all open protein models that they used.
What is not known
The authors did not see the training data. They use the results of the models. Anthropic gave no information. The result does not include DNA: Opus 5.5 predicts human and mouse coding sequences, but not the remaining genome.
Insight: Tests from outside can show what a vendor used
In October 2026, Anthropic said that tests of a model in use show more than a check of its training data. These authors made such a test from external access. It reads as if result tests can show the data of a vendor, and the vendor does not say it. Buyers can use the same method.
Sources
- Claude was trained on protein sequences at scalexinmingtu.cn
Posted