Claude optimized 36 biological models in under 4 weeks, achieving an average speedup of about 4x

Beating AI News Flash: Anthropic had Claude perform a round of performance optimization on 36 open-source biological models. In under 4 weeks, these models ran about 4 times faster on average, covering tasks such as protein structure prediction, protein design, and genomic analysis. All related optimization code has been fully open-sourced.

What Claude did was not just modify Python code; it also wrote GPU kernels such as FlashPairformer itself, directly improving GPU computational efficiency from the ground up. In official tests, some originally very time-consuming computations became up to 3.2 times faster. When encountering different models, Claude would also identify bottlenecks one by one, remove redundant computations, or cache frequently used data in advance.

The entire process was supervised by two Anthropic biological modeling researchers, but neither had prior experience in inference optimization or kernel engineering. Even when requiring results to be exactly identical before and after optimization, these models still achieved an average speedup of about 1.6 times; if allowing minor differences within the normal range in model results, the average speedup could reach about 4 times.

In a previous protein design experiment by Anthropic, each target required up to about 2,500 H100 GPU hours. After optimization, a single H200 running for 24 hours could achieve a similar effect, with the combined cost of GPU plus Claude at about $150. Experiments that originally required burning thousands of GPU hours can now be completed in about a day.