Independent researcher Xinming Tu and co-author Leo Tianlai Chen argue that Anthropic’s latest models absorbed protein sequences in bulk during training, based on a battery of tests published on Tu’s blog, xinmingtu.cn. The conclusion is an inference from outside the lab. Anthropic has not described what went into Claude Opus 5.5 or Claude Fable 5.2, and the authors cannot see the training data.
The test is simple to state. A model receives the opening stretch of a protein, written as a string of letters where each letter is an amino acid (the small units that proteins are built from), and must write the next ten. Only the first is scored. The authors then repeat the question on a shuffled copy of the same protein. Shuffling keeps the same letters but scrambles their order, so a model that only knows the usual letter frequencies earns the same score on both. Any gap between the two versions measures knowledge of real sequences.
The proteins were chosen to rule out memorization from papers. Tu and Chen drew 668 proteins from bacteria and archaea, spread over 167 Pfam families (catalogued groups of related proteins) labeled as having unknown function. No family has a solved 3D structure or more than three papers, and no protein has a close relative in UniProt, the main public sequence database. Each protein was cut at five points, and models ran at their lowest thinking setting.
On that set, Opus 5.5 got the next amino acid right 34.9 percent of the time on real proteins and 7.5 percent on shuffled ones, a gap of 27.4 points. Fable 5.2 scored 54 percent against 9 percent, a gap of 45.1 points. Across the 20 language models tested, the authors report that GPT, Gemini and seven 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) stayed near zero. The gap first shows up in Opus 4.7, grows with Opus 5.5, and grows again with Fable 5.2.
The natural yardstick is a protein language model, a network trained only on sequences. ProGen3-3B scored a gap of 32.5 points and ProtGPT3-10B, the largest open model of its kind, 27.7. Fable 5.2 beat both by a wide margin, while Opus 5.5 matched ProtGPT3-10B and trailed ProGen3-3B. Three ESM models, which normally read a residue from both directions, landed between 7.0 and 13.0 points when forced to read left to right.
Tu and Chen then asked whether training data explains the pattern. Bigger protein families should be easier if models learned from sequence databases, and Opus 5.5 rose with family size alongside the two protein language models. Papers did not explain it: across the 128 families with no papers at all, the gaps were 45.9 points for Fable 5.2 and 26.8 for Opus 5.5. Reasoning did not either. The API reported thinking tokens on 9 percent of Fable 5.2’s answers and 36 percent of Opus 5.5’s, and answers without thinking were just as accurate. GPT-6 Astra and GPT-6.1 Sol reason at length on each question, yet still stayed near zero.
The skill is narrow. Opus 5.5 predicted human and mouse coding DNA but not the rest of the genome. It wrote all 20 nucleotides right in 25 of 500 human coding windows, where a blind guess would manage that about once in a trillion.
Several limits apply. This is a blog post from a small team, and the evidence is behavioral: it shows what the models can do, not what they saw. The Fable models ran through OpenRouter, and a safety filter at Anthropic rejected 44 percent of the questions sent to the older Claude Fable 5, so the authors left that model out of the main figures. The full run cost about $730 in API fees, and the code, test sets and outputs sit in a repository the authors call llm-bioseq-probe, so others can rerun it.
Labs building biology products on Claude should treat the result as a reason to test on their own obscure sequences, since a gap this size is cheap to check. The authors also leave open whether Claude could fold a protein or rank mutations the way protein language models do.
Xinming Tu and Leo Tianlai Chen, “Claude was trained on protein sequences at scale,” xinmingtu.cn, November 2026.