Claude independently wrote GPU code and sped up more than 30 open-source biomolecular neural networks by an average factor of 4
Claude independently wrote GPU code and sped up more than 30 open-source biomolecular neural networks by an average factor of 4
On September 17, Anthropic published a technical report: over the course of a month, its research model Claude independently optimized the code of 36 versions across more than 30 open-source scientific programs that predict three-dimensional protein structures, design new molecules, and analyze genomic data. The work was led by two staff members with biology backgrounds but no GPU programming experience. Speedups and accuracy were validated across thousands of predictions. All code is open-source.
Neural networks such as AlphaFold3 and its open-source counterparts spend most of their computation on operations that compare triplets of molecular segments to determine the molecule's spatial arrangement: doubling the molecule's size increases computational cost eightfold. The standard way to speed this up is to rewrite low-level GPU code, known as kernels. Claude wrote such kernels at Anthropic's direction. The resulting FlashPairformer kernel set outperformed NVIDIA's standard code by 1.7–3.2x on one of these operations and by 2.7–2.9x on another.
Beyond these kernels, Claude applied the same approach to the full set of 36 program versions. An exact-match mode, preserving the original output while running faster, yielded an average 1.6x speedup; a mode allowing a small accuracy tradeoff achieved an average 4.1x speedup. Across nearly two thousand model-molecule test pairs, the share of accurate predictions was virtually unchanged: 54.8% for the original code versus 54.2–55% for the accelerated versions, a difference statistically indistinguishable from noise.
A separate mode removed the GPU memory constraint and for the first time allowed existing open-source models on a single server to accurately predict the structure of enormous molecular machines such as the E. coli ribosome, even though these models were trained on fragments at least 13 times smaller. When Claude was asked to predict entire viral capsids 40–90 times larger than the training data, all attempts ran to completion, but none produced a correct structure: the output was a collapsed globule roughly a quarter of the true size.
The technical report containing these figures was written by Claude itself, under human supervision.
The work was prompted by an earlier campaign: in August, Claude designed 1,320 binder proteins for 15 targets, and laboratories confirmed 354 of them as binding, but a single target cost up to $10,000 and roughly 2,500 GPU hours, a sum out of reach for most laboratories. A pilot experiment had validated the approach beforehand: Claude Mythos 5.1 manually sped up 7 models by up to 2.5x with no loss in accuracy. Three versions of Claude then replicated the same work across all accelerated models without agents or human involvement, on a single GPU in 24 hours and at a cost of approximately $150; the new variants have so far been validated only computationally. At the same time, Anthropic and the protein design laboratory Adaptyv Bio announced a competition with a prize pool of up to $1 million in Claude credits.