← All stories
● Covered by 1 source · 1 reportMedium impact1 neutral

Claude Models Show Evidence of Training on Protein Sequences at Scale

🔄 Updated 3d ago
New to BrevFeed? We gather this story from every outlet covering it into one summary — ranked by real-world impact, not just the latest headline — so you never miss what matters. What is BrevFeed? →

Key points

  • Claude Fable 5.2 and Opus 5.5 predict next amino acids in novel proteins.
  • Other frontier LLMs (GPT, Gemini) do not show this capability.
  • Claude models perform better on larger protein families, like protein language models.
  • Fable 5.2 outperforms open protein language models; Opus 5.5 matches ProtGPT3-10B.

Claude's Protein Prediction Capability

Recent testing indicates that Anthropic's latest Claude models, specifically Fable 5.2 and Opus 5.5, possess the ability to accurately predict the next amino acid in previously unstudied bacterial proteins. This capability is demonstrated by their performance on a 'next-amino-acid' test, where they predict sequences from real proteins but fail when presented with shuffled versions of the same proteins.

Distinction from Other LLMs

Unlike Claude, other leading large language models such as GPT, Gemini, and several open-weight models (DeepSeek V4.1 Flash, Qwen 3.8, Kimi K3.1 Pro, GLM-5.3, MiniMax M3, gpt-oss-120b, and Nemotron 3 Ultra) do not exhibit this protein sequence prediction ability. These other models show near-zero accuracy in the same tests, even when continuing raw FASTA text.

Performance Metrics and Comparison

Fable 5.2 correctly predicts the next amino acid 54% of the time on real proteins, compared to 9% on shuffled sequences. This performance gap is consistent across most protein families tested. The capability appeared in Opus 4.7, increased in Opus 5.5, and further improved in Fable 5.2. Fable 5.2 surpasses all tested open protein language models, while Opus 5.5 performs on par with ProtGPT3-10B, one of the largest specialized protein models.

Testing Methodology

The test set comprised 668 bacterial and archaeal proteins from 167 Pfam families with unknown functions. These families had no solved structures and fewer than four associated papers, and no protein had a close relative in UniProt. Models were given the start of a protein sequence and asked to predict the next 10 amino acids, with only the first prediction scored. A shuffled version of each protein was also presented to isolate knowledge of sequence order.

Implications for AI Training

The observed capabilities suggest that the latest Claude models were trained on extensive protein sequence data, integrating biological sequence understanding into their general language model architecture. This indicates a broader scope of training data for these models compared to their contemporaries.

✨ This summary was generated by AI from the outlets' reporting listed below. It is not independently verified and may contain errors — check the original sources. How BrevFeed works →

The daily brief

One email each morning: the day's tech stories, clustered across outlets and summarized. No account needed.

One email a day. Unsubscribe in one click, any time.

Today's brief

Spend a few minutes, get the whole day. Every topic's top stories in one hands-free rundown — listen, watch, or read the transcript.

~4 min · 3 stories · Nov 21

▶ Play today's brief Listen on Spotify

New every morning, and the back catalogue is archived by date.

Reporting from

The latest Claude models, Fable 5.2 and Opus 5.5, demonstrate the ability to predict subsequent amino acids in little-studied bacterial proteins, a capability not observed in other frontier large language models. This suggests these Claude models were trained on protein sequences at scale, performing comparably to specialized protein language models.