Anthropic has published an account of using Claude to optimise more than 30 open-source deep learning models used in biology — structure prediction, protein design, genomics and protein language models — in just under four weeks, supervised by two members of its technical staff.
The gains reported are in compute, not accuracy. The average speed-up was roughly 4x with minimal loss of precision, or nearly 2x with identical outputs. New attention kernels, which Anthropic calls FlashPairformer, beat field standards by 2.7–2.9x on triangle attention and 1.7–3.2x on triangle multiplication. A memory mode lets a single Nvidia GPU predict structures above 10,000 tokens, and up to 70,000 — large enough for the bacterial ribosome, the proteasome and human mitochondrial complex I. In a protein design run, comparable results in silico came from two orders of magnitude fewer GPU hours, at about $150 of combined GPU and token cost.
Much of structural biology is rationed by who can afford the GPUs, and these are engineering optimisations applied to models that are already public. The caveats: every figure here is Anthropic’s own, the protein design results are in silico rather than wet-lab validated, and Anthropic is running a related competition with Adaptyv Bio offering up to $1 million in Claude credits — so it is not a disinterested party.