Four weeks, thirty-odd models
Anthropic published work on Thursday in which Claude, running inside the company’s Claude Science setup, rewrote more than 30 open-source deep learning models used in biology so that they run faster. The company says the average speedup across structure prediction and design models was roughly four times, and that two Anthropic staff with no prior experience in inference optimisation oversaw the work over just under four weeks.
The output is 36 optimised packages across six families of model: 14 for co-folding and structure prediction, seven for genomics, six for structure generation, and three each for inverse folding, hallucination and protein language models.
What got faster, and by how much
Anthropic reports three modes. An exact mode that produces identical outputs runs about 1.6 times faster. A fast mode across 13 models runs about 4.1 times faster. A low-memory mode runs about 3.4 times faster. Underneath sits a custom set of GPU kernels the company calls FlashPairformer, which it says beats Nvidia’s own cuEquivariance implementations: 2.7 to 2.9 times faster on triangle attention, and 1.7 to 3.2 times faster on triangle multiplication.
Every one of those numbers is Anthropic’s own measurement, and all of them were taken on Nvidia H100 GPUs. No outside group has reproduced them.
Bigger molecules on a single node
The practical change is memory. The low-memory mode lets one Nvidia GPU node predict biomolecular systems of more than 10,000 tokens — amino acids, nucleotides and the atoms of small molecules and ions — and Anthropic says it ran inference on systems above 70,000. It reports folding human mitochondrial complex I, the TRiC chaperone complex, proteasomes and bacterial ribosomes.
It also describes a protein design campaign that reached comparable binding scores to earlier work using roughly two orders of magnitude fewer GPU hours, at about $150 of combined GPU and token cost. Anthropic states plainly that those designs were never tested in a laboratory and that the scores are computational.
What did not work
The write-up records the failures too. Predictions of large viral capsids collapsed into a compact ball with very low accuracy scores, Unite.AI noted. The repository is published as a reference release the company says it will not maintain.
A competition to settle it at a bench
Anthropic has open-sourced the optimised code and is running a protein design competition with Adaptyv Bio on five problems at the edge of current capability, among them species cross-reactivity, pH sensitivity, peptide-MHC specificity and G protein-coupled receptor targets. The pool is up to $1m in Claude credits and $250,000 in Modal compute credits, with DNA synthesis from Twist Bioscience and wet lab validation for more than 5,000 designs. Unite.AI reports per-team caps of $50,000 for academic entrants and $25,000 for industry ones.
What to watch
Applications close on 24 September and results are due on 15 December. That second date is the one that matters. Until designs come back from a bench, the entire claim rests on scores a computer gave itself.
