What was timed is duration
On 17 September Anthropic wrote that an internal research model inside Claude Science optimized more than 30 open-source biomolecular models in under four weeks, speeding them up by roughly 4 times on average on the company's own measurements. A low-memory mode is said to allow prediction of systems larger than 10,000 tokens on a single Nvidia GPU node. Tokens here count amino acids, nucleotides and atoms from small molecules. The speed-up is tied to geometry operations such as triangle attention and triangle multiplication.[1]
The claim is not a patient outcome or a binding constant. The post also recalls that Mythos 5.1 had sped seven open biology models by up to 2.5 times, again on company measurement. Anthropic says it will open-source the optimized code and, with Adaptyv Bio, run a contest promising wet-lab validation for more than 5,000 designs and up to 1 million dollars in Claude credits. The prize and the 4-times figure are vendor claims; there is no independent benchmark table.[1]
The next evidence is wet-lab
Claude had earlier been described as designing de novo binders by orchestrating open protein-design and structure models; that too is an orchestration claim. Speed makes the same task cheaper; it does not say whether a binder holds. In the 17 September Mendelian column, a hyperbolic embedding score still stopped short of a patient decision. A wall-clock here likewise does not carry the laboratory meaning of a bond. A fair alternative is that faster code simply yields more candidates at the same accuracy, and the contest will show it. That needs the code, a named benchmark and a wet-lab share to appear together.[1], [2]