Anthropic published results yesterday that will make somebody in structural biology uneasy: an internal research model optimised more than 30 open-source models for biomolecular tasks. On average they run about 4x faster with minimal loss of precision, and nearly 2x faster with identical outputs. The code is open and so is the technical report.
FlashPairformer
Modern structure prediction — AlphaFold3, OpenFold3, Boltz-2 — spends most of its runtime in two operations: triangle attention and triangle multiplication. Both are cubic in time and memory. Double the system and you pay 8x; triple it and you pay 27x.
The standard answer is kernels, the low-level translation layers for GPUs. NVIDIA did this with cuEquivariance and more recently with its BioNeMo Inference Runtime. Claude built FlashPairformer out of that work and beats the field standard by 2.7–2.9x on triangle attention and 1.7–3.2x on triangle multiplication, depending on configuration.
Big mode: a ribosome on one node
The second half is memory. Large molecular machines — ribosomes, respiratory complexes, chaperones — are built from dozens of components, and their function depends on how those fit together. Predicting systems that size has meant spreading inference across multiple GPU nodes.
Claude’s low-memory “Big” mode handles systems over 10,000 tokens on a single NVIDIA node. Successfully folded structures include human mitochondrial complex I, the TRiC chaperone complex, a proteasome and a bacterial ribosome, each close to its experimentally determined structure. For scale: AlphaFold3 predicted the 40S ribosome at 7,663 tokens.
They also pushed past the edge. On a single 8-GPU B300 node, Claude ran predictions of entire viral capsids from 31,000 to over 70,000 tokens. Those collapse — nearly two orders of magnitude beyond the training context, the models don’t generalise. But the run completes at all, on one node.
$10,000 becomes $150
In the earlier work, Claude got a prompt of roughly 16,000 words, could use sub-agents, and was allowed up to $10,000 per target on Modal — about 2,500 H100 hours inside 24 hours.
This time: a single model, one H200, 24 hours of wall time, a prompt of about 1,100 words and a reference sheet for the pre-installed tools. No sub-agents, nobody steering the designs. Averaged over 16 targets, Mythos 5.1, Mythos 5 and Opus 5 reach roughly the same ipSAE scores as the old campaign — at about two orders of magnitude fewer GPU hours, and a combined spend of around $150 on compute and tokens.
There’s a competition attached, co-sponsored with Adaptyv Bio: five problems, up to $1 million in Claude credits, $250,000 in Modal compute, and wet lab validation for more than 5,000 designs.
Two supervisors with no kernel experience is the real number
One clause in the post says more than any speed-up figure: the work was supervised by two Anthropic staff who know biomolecular modelling and had no prior experience with inference optimisation or kernel engineering.
Optimisations like these take an experienced team weeks per model, and they usually don’t transfer to the next one. Here it was thirty-odd models in just under four weeks. That squares with the figure Anthropic put out this morning: Claude now carries out 26 percent of its own AI research.
Those 5,000 lab validations from the competition end up in real labs, which fits neatly with Anthropic having just built a wet lab of its own. The pipette is still the last step.
Sources: