Signum
Feed
Useful signal18 Nov 2026high confidence

Independent test finds latest Claude models (Fable 5.2, Opus 5.5) predict next amino acids in little-studied proteins, indicating training on protein sequences; other frontier models show no such ability

Edited by · written with AI from the linked sources (how)

A research write-up (updated 18 Nov 2026) reports a next-amino-acid test on 668 bacterial/archaeal proteins from 167 unknown-function Pfam families. Real-minus-shuffled accuracy is +45.1 points for Claude Fable 5.2 and +27.4 for Opus 5.5. Fable 5.2 beats every open protein language model tested, including ProGen3-3B (+32.5). Opus 5.5 is level with ProtGPT3-10B (+27.7). GPT, Gemini and open-weight LLMs stay near zero. The update adds 56 proteins published in Aug 2026, after the training cutoff, which the latest Claude models still predict, plus more PLMs and GPT-6.2 models. Performance scales with family size and holds in families with no papers. This is inference from black-box testing, not a confirmation by Anthropic.

Capability

Entities: Anthropic, Claude Fable 5.2, Claude Opus 5.5, Claude Fable 5.1, GPT-6.2 Sol, GPT-6.2 Astra

66Useful signal
1 source
0 primary
Was this useful?
01

What happened

An independent write-up, updated 18 Nov 2026, tested whether language models can predict the next amino acid in 668 bacterial and archaeal proteins from 167 families with no known function. Scored as real minus shuffled sequence accuracy, Claude Fable 5.2 gains 45.1 points and Opus 5.5 gains 27.4. Fable 5.2 beats every open protein language model tested, including ProGen3-3B (+32.5), and Opus 5.5 roughly matches ProtGPT3-10B (+27.7). GPT, Gemini and open-weight general LLMs stay near zero. The update adds 56 proteins published in Aug 2026, after the stated training cutoff, which the newest Claude models still predict. The authors infer that Claude was trained on raw protein sequences. Anthropic has not confirmed this.

02

Why it matters

If the inference holds, a general-purpose chat model has picked up protein-language-model skills that no other frontier LLM tested shows. That matters to biology researchers and to developers weighing whether one general model could replace or complement specialist tools. It also hints at a data-curation difference between labs. It does not yet change anything in deployment. The test measures next-residue guessing, not useful tasks such as structure, function or design. Nobody should switch tools based on it without task-level benchmarks.

03

What is noise

The headline states as fact that Claude "was trained on" protein sequences, but this is a black-box inference and Anthropic has not said so. Next-amino-acid accuracy is a proxy and says little about whether the models are useful for real protein work. The post-cutoff test is reassuring, but it covers only 56 proteins, and homologous sequences may already be in training data. No primary links are supplied, and this is not peer reviewed.

04

Watch next

  1. 01Any statement from Anthropic, such as a model card or data disclosure, confirming or denying protein sequence training.
  2. 02Independent replication by other groups, ideally using larger post-cutoff protein sets and benchmarks on real tasks such as function prediction or variant effects, not only next-residue accuracy.
  3. 03Whether other labs, especially OpenAI and Google, retest or release models that close the gap, and whether the +45 point lead of Fable 5.2 over open protein language models holds with new proteins.

Coverage

1 story

Newsletter

Get the no-hype AI news straight to your inbox

Weekly. No spam, unsubscribe any time. Privacy

More capability signals

Full feed →